Huang - Machine Learning Modeling Data Locally and Globally - 2008

ADVANCED TOPICS
IN SCIENCE AND TECHNOLOGY IN CHINA
ADVANCED TOPICS
IN SCIENCE AND TECHNOLOGY IN CHINA
Zhejiang University is one of the leading universities in China. In Advanced
Topics in Science and Technology in China, Zhejiang University Press and
Springer jointly publish monographs by Chinese scholars and professors, as
well as invited authors and editors from abroad who are outstanding experts
and scholars in their elds. This series will be of interest to researchers, lecturers, and graduate students alike.
Advanced Topics in Science and Technology in China aims to present the
latest and most cutting-edge theories, techniques, and methodologies in various research areas in China. It covers all disciplines in the elds of natural
science and technology, including but not limited to, computer science, materials science, life sciences, engineering, environmental sciences, mathematics,
and physics.
Kaizhu Huang
Haiqin Yang
Irwin King
Michael Lyu
Machine Learning
Modeling Data Locally and Globally
With 53 gures
AUTHORS:
Dr. Kaizhu Huang,
Dept. of CSE,
Chinese Univ. of Hong Kong,
Shatin. N.T. HK,
China
Email: kzhuang@cse.cuhk.edu.hk
Dr. Haiqin Yang,

Dept. of CSE,
Shatin. N.T. HK,
China
Email:hqyang@cse.cuhk.edu.hk
Prof. Irwin King,

Dept. of CSE,
Shatin. N.T. HK,
China
Email: king@cse.cuhk.edu.hk
Dr. Michael Lyu,

Dept. of CSE,
Shatin. N.T. HK,
China
Email:lyu@cse.cuhk.edu.hk
ISBN 978-7-308-05831-5 Zhejiang University Press, Hangzhou

ISBN 978-3-540-79451-6 Springer Berlin Heidelberg New York
e-ISBN 978-3-540-79452-3 Springer Berlin Heidelberg New York
Series ISSN 1995-6819 Advanced topics in science and technology in China
Series e-ISSN 1995-6827 Advanced topics in science and technology in China
Library of Congress Control Number : 2008925536
This work is subject to copyright. All rights are reserved, whether the whole or part
of the material is concerned, specically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microlm or in any other
way, and storage in data banks. Duplication of this publication or parts thereof is
permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from
Springer-Verlag. Violations are liable to prosecution under the German Copyright
Law.
c 2008 Zhejiang University Press, Hangzhou and Springer-Verlag GmbH Berlin

Heidelberg
Co-published by Zhejiang University Press, Hangzhou and SpringerVerlag GmbH Berlin Heidelberg
Springer is a part of Springer Science+Business Media
springer.com
The use of general descriptive names, registered names, trademarks, etc. in this
publication does not imply, even in the absence of a specic statement, that such
names are exempt from the relevant protective laws and regulations and therefore
free for general use.
Cover design: Joe Piliero, Springer Science + Business Media LLC, New York
Printed on acid-free paper
Preface
Machine Learning: Modeling Data Locally and Globally delivers the

main contemporary themes and tools in machine learning including probabilistic generative models and Support Vector Machines. These themes are
discussed or reformulated from either a local view or a global view. Dierent from previous books that only investigate machine learning algorithms
locally or globally, this book presents a unied and new picture for machine
learning both locally and globally. Within the new picture, various seemly
dierent machine learning models and theories are bridged in an elegant and
systematic manner. For precise and thorough understanding, this book also
presents applications of the new hybrid theory.
This book not only provides researchers with the latest research results
lively and timely, but also presents an excellent overview on machine learning.
Importantly, the new line of learning both locally and globally goes through
the whole book and makes various learning models understandable to a large
proportion of audience including researchers in machine learning, practitioners in pattern recognition, and graduate students.
The Chinese Univ. of Hong Kong,
Jan. 2008
Kaizhu Huang
Haiqin Yang
Irwin King
Michael R. Lyu
Contents
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Learning and Global Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Learning and Local Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Hybrid Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Major Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 Book Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
1
3
5
5
8
8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Global Learning vs. Local Learning . . . . . . . . . . . . . . . . . . . . . . .

2.1 Problem Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Global Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Generative Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.2 Non-parametric Learning . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.3 The Minimum Error Minimax Probability Machine . . .
2.3 Local Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Hybrid Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Maxi-Min Margin Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
15
16
16
19
21
22
23
24
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3
A General Global Learning Model: MEMPM . . . . . . . . . . . . .

3.1 Marshall and Olkin Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Minimum Error Minimax Probability Decision Hyperplane . . .
3.2.1 Problem Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.3 Special Case for Biased Classications . . . . . . . . . . . . . .
3.2.4 Solving the MEMPM Optimization Problem . . . . . . . . .
3.2.5 When the Worst-case Bayes Optimal Hyperplane
Becomes the True One . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
30
31
31
32
33
34
39
VIII
Contents
3.2.6 Geometrical Interpretation . . . . . . . . . . . . . . . . . . . . . . . .

3.3 Robust Version . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Kernelization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.1 Kernelization Theory for BMPM . . . . . . . . . . . . . . . . . . .
3.4.2 Notations in Kernelization Theorem of BMPM . . . . . . .
3.4.3 Kernelization Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.1 Model Illustration on a Synthetic Dataset . . . . . . . . . . .
3.5.2 Evaluations on Benchmark Datasets . . . . . . . . . . . . . . . .
3.5.3 Evaluations of BMPM on Heart-disease Dataset . . . . . .
3.6 How Tight Is the Bound? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.7 On the Concavity of MEMPM . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.8 Limitations and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
45
46
47
48
49
50
50
50
55
56
60
65
66
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4
Learning Locally and Globally: Maxi-Min Margin Machine

4.1 Maxi-Min Margin Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Separable Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.2 Connections with Other Models . . . . . . . . . . . . . . . . . . . .
4.1.3 Nonseparable Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.4 Further Connection with Minimum Error Minimax
Probability Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Bound on the Error Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4 Kernelization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Foundation of Kernelization for M4 . . . . . . . . . . . . . . . . .
4.4.2 Kernelization Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.1 Evaluations on Three Synthetic Toy Datasets . . . . . . . .
4.5.2 Evaluations on Benchmark Datasets . . . . . . . . . . . . . . . .
4.6 Discussions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
71
71
74
78
80
82
84
85
85
86
88
88
90
93
93
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5
Extension I: BMPM for Imbalanced Learning . . . . . . . . . . . .

5.1 Introduction to Imbalanced Learning . . . . . . . . . . . . . . . . . . . . . .
5.2 Biased Minimax Probability Machine . . . . . . . . . . . . . . . . . . . . .
5.3 Learning from Imbalanced Data by Using BMPM . . . . . . . . . .
5.3.1 Four Criteria to Evaluate Learning from Imbalanced
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2 BMPM for Maximizing the Sum of the Accuracies . . . .
5.3.3 BMPM for ROC Analysis . . . . . . . . . . . . . . . . . . . . . . . . .
97
98
98
100
100
101
102
Contents
5.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5.4.1 A Toy Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.2 Evaluations on Real World Imbalanced Datasets . . . . .
5.4.3 Evaluations on Disease Datasets . . . . . . . . . . . . . . . . . . . .
5.5 When the Cost for Each Class Is Known . . . . . . . . . . . . . . . . . .
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IX
102
102
104
111
114
115
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6
Extension II: A Regression Model from M4 . . . . . . . . . . . . . . .

6.1 A Local Support Vector Regression Model . . . . . . . . . . . . . . . . .
6.1.1 Problem and Model Denition . . . . . . . . . . . . . . . . . . . . .
6.1.2 Interpretations and Appealing Properties . . . . . . . . . . . .
6.2 Connection with Support Vector Regression . . . . . . . . . . . . . . . .
6.3 Link with Maxi-Min Margin Machine . . . . . . . . . . . . . . . . . . . . .
6.4 Optimization Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Kernelization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.6 Additional Interpretation on wT i w . . . . . . . . . . . . . . . . . . . . .
6.7 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.7.1 Evaluations on Synthetic Sinc Data . . . . . . . . . . . . . . . . .
6.7.2 Evaluations on Real Financial Data . . . . . . . . . . . . . . . . .
6.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
119
121
121
122
122
124
124
125
127
128
128
130
131
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7
Extension III: Variational Margin Settings within Local

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1 Support Vector Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Problem in Margin Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3 General -insensitive Loss Function . . . . . . . . . . . . . . . . . . . . . . .
7.4 Non-xed Margin Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4.1 Momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4.2 GARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5.1 Accuracy Metrics and Risk Measurement . . . . . . . . . . . .
7.5.2 Momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5.3 GARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.6 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
133
134
136
136
139
139
140
141
141
142
149
155
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8
Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

8.1 Review of the Journey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.1 Inside the Proposed Models . . . . . . . . . . . . . . . . . . . . . . . .
161
161
163
163
Contents
8.2.2 Beyond the Proposed Models . . . . . . . . . . . . . . . . . . . . . . 164

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
1
Introduction
The objective of this book is to establish a framework which combines two

dierent paradigms in machine learning: global learning and local learning.
The combined model demonstrates that a hybrid learning of these two different schools of approaches can outperform each isolated approach both
theoretically and empirically. Global learning focuses on describing a phenomenon or modeling data in a global way. For example, a distribution over
the variables is usually estimated for summarizing the data. Its output can
usually reconstruct the data. This school of approaches, including Bayesian
Networks [8, 13, 30], Gaussian Mixture Models [3, 21], and Hidden Markov
Models [2, 25], has a long and distinguished history, which has been extensively applied in articial intelligence [26], pattern recognition [9], and computer vision [7]. On the other hand, local learning does not intend to summarize a phenomenon, but builds learning systems by concentrating on some
local parts of data. It lacks the exibility yet surprisingly demonstrates superior performance to global learning according to recent researches [4, 16, 15].
In this book, a bridge has been established between these two dierent
paradigms. Moreover, the resulting principled framework subsumes several
important models, which respectively locate themselves into the global learning paradigm and the local learning paradigm.
In this chapter, we address the motivations of the two dierent learning
frameworks. As a summary, we present the objectives of this book and outline
the main models or the contributions. Finally, we provide an overview of the
rest of this book.
1.1 Learning and Global Modeling

When studying real world phenomena, scientists are always wondering whether
some underlying laws or nice mathematical formulae exist for governing these
complex phenomena. Moreover, in practice, due to incomplete information,
1 Introduction
the phenomena are usually nondeterministic. This motivates to base probabilistic or statistical models to perform a global investigation on sampled data
from the phenomena. A common way for achieving this goal is to t a density
on the observations of data. With the learned density, people can then incorporate prior knowledge, conduct predictions, and perform inferences and
marginalizations. One main category in the framework of global learning is
the so-called generative learning. By assuming a specic mathematical model
on the observations of data, e.g. a Gaussian distribution, the phenomena can
therefore be described or re-generated. Fig. 1.1 illustrates such an example.
In this gure, two classes of data are plotted as s for the rst class and
s for the other class. The data can thus be modeled as two dierent mixtures of Gaussian distributions as illustrated in Fig. 1.2. By knowing only the
parameters of these distributions, one can then summarize the phenomena.
Furthermore, one can clearly employ this information to distinguish one class
of data from the other class or simply know how to separate two classes. This
is also well-known as Bayes optimal decision problems [12, 6].
Fig. 1.1. Two classes of two-dimensional data
In the development of learning approaches within the community of machine learning, there has been a migration from the early rule-based methods [11, 32] wanting more involvement of domain experts, to widely-used
probabilistic global models mainly driven by data itself [5, 9, 14, 17, 22, 33].
However, one question for most probabilistic global models is what kind of
global models, or more specically, which type of densities should be specied beforehand for summarizing the phenomena. For some tasks, this can be
prescribed by a slight introduction of domain knowledge from experts. Unfortunately, due to both the increasing sophistication of the real world learning
tasks and active interactions among dierent subjects of research, it is more
1.2 Learning and Local Modeling
Fig. 1.2. An illustration of distribution-based classications (also known as

the Bayes optimal decision theory). Two Gaussian mixtures are engaged to
model the distribution of two classes of data respectively. The distribution
can then be used to construct the decision plane
and more dicult to obtain fast and valuable suggestions from experts. A further question is thus proposed, i.e. what is the next step in the community
of machine learning, after experiencing a migration from rule-based models
to probabilistic global models? Recent progress in machine learning seems to
imply local learning as a solution.
1.2 Learning and Local Modeling

Global modeling addresses describing phenomena, no matter whether the
summarized information from the observations is applicable to specic tasks
or not. Moreover, the hidden principle under global learning is that information can be accurately extracted from data. On the other hand, local
learning [10, 27, 28] which recently attracts active attention in the machine
learning community, usually regards that a general and accurate global learning is an impossible mission. Therefore, local learning focuses on capturing
only local yet useful information from data. Furthermore, recent research
progress and empirical study demonstrate that this much dierent learning
paradigm is superior to global learning in many facets.
In further details, instead of globally modeling data, local learning is more
task-oriented. It does not aim to estimate a density from data as in global
learning, which is usually an intermediate step for many tasks such as pattern
recognitions (note that the distribution or density obtained by global learning actually is not directly related to the classication itself); it also does not
intend to build an accurate model to t the observations of data globally. Differently, it only extracts useful information from data and directly optimizes
the learning goal. For example, when used in learning classiers from data,
only those observations of data around the separating plane need to be accurate, while inaccurate modeling over other data is certainly acceptable for
1 Introduction
the classication purpose. Fig. 1.3 illustrates such a problem. In this gure,
the decision boundary is constructed only based on those lled points, while
other points make no contributions to the classication plane (the decision
boundary is given based on the Gabriel Graph method [1, 18, 34]).
Fig. 1.3. An illustration of local learning (also known as the Gabriel

Graph classication). The decision boundary is just determined by
some local points indicated as lled points
However, although containing promising performance, local learning appears to locate itself at another extreme end to global learning. Employing
only local information may lose the global view of data. Consequently, sometimes, it cannot grasp the data trend, which is critical for guaranteeing better
performance for future data. This can be seen in the example as illustrated
in Fig. 1.4. In this gure, the decision boundary (also constructed by the
Gabriel Graph classication) is still determined by some local points indicated as lled points. Clearly, this boundary does not grasp the data trend.
Fig. 1.4. An illustration on that local learning cannot grasp data trend.
The decision boundary (constructed by the Gabriel Graph classication)
is determined by some local points indicated as lled points. It, however,
loses the data trend. The decision plane should be obviously closer to the
lled squares rather than locating itself in the middle of lled s and s
1.4 Major Contributions
More specically, the class associated with s is obviously more scattered

than the class
associated with s on the axis indicated as dashed line. Therefore, a
more promising decision boundary should lie closer to lled s than those
lled s instead of lying midway between lled points. A similar example
can also be seen in Chapter 2 on a more principled local learning model, i.e.
the current state-of-the-art classier, Support Vector Machines (SVM) [31].
Targeting this problem, we then suggest a hybrid learning in this book.
1.3 Hybrid Learning

There are complementary advantages for both local learning and global learning. Global learning summarizes data and provides practitioners with knowledge on the structure, independence, and trend of data, since with the precise
modeling of phenomena, the observations can be accurately regenerated and
therefore can be studied or analyzed thoroughly. However, this also presents
diculties in how to choose a valid model to describe all the information
(also called the problem of model selection). In comparison, local learning
directly employs part of information, critical for the specic oriented tasks,
and does not assume models to re-synthesize/restore the whole road-map of
data. Although demonstrated to be superior to global learning in many facets
of machine learning, it may lose some important global information. The
question here is thus, can reliable global information, independent of specic
model assumptions, be combined into local learning? This question clearly
motivates a hybrid learning of two largely dierent schools of approaches,
which is also the focus of this book.

In this book, we aim to describe a hybrid learning scheme to combine two
dierent paradigms, namely global learning and local learning. Within this
scheme, we propose a hybrid model, named the Maxi-Min Margin Machine
(M4 ), demonstrated to contain both the merits of global learning in representing data and the advantages of local learning in handling tasks directly
and eectively. Moreover, adopting the viewpoint of local learning, we also
introduce a global learning model, called the Minimum Error Minimax Probability Machine (MEMPM), which does not assume specic distributions on
data and thus distinguishes itself from traditional global learning approaches.
The main models discussed in this book are briey described as follows.
The Maxi-Min Margin Machine model, a hybrid learning framework successfully combining global learning and local learning
1 Introduction
A unied framework of many important models

As will be demonstrated, our proposed hybrid model successfully unies both important models in local learning, e.g. the Support Vector
Machines [4], and signicant models in global learning, such as the
Minimax Probability Machine (MPM) [19] and the Fisher Discriminant Analysis (FDA) [9].
With the generalization Guarantee
Various statements from many views such as the sparsity and Marshall and Olkin Theory [20, 23] will be presented for providing the
generalization bound for the combined approach.
A sequential Conic Programming solving method
Besides the theoretic advantages of the proposed hybrid learning, we
also tailor a sequential Conic Programming method [24, 29] to solve
the corresponding optimization problem. The computational cost is
shown to be polynomial and thus the proposed M4 model can be
solved practically.
The Minimum Error Minimax Probability Machine, a general global
learning model
A worst-case distribution-free Bayes optimal classier
Dierent from traditional Bayes optimal classiers, MEMPM does
not assume distributions for the data. Starting with the Marshall
and Olkin theory, this model attempts to model data under the minimax schemes. It does not intend to extract exact information but the
worst-case information from data and thus presents an important
progress in global learning.
Derive an explicit error bound for future data
Inheriting the advantages of global learning, the proposed general
global learning method contains an explicit worst-case error bound
for future data under a mild condition. Moreover, the experimental
results suggest that this bound is reliable and accurate.
Propose a sequential Fractional Programming optimization
We have proposed a Fractional Programming optimization method
for the MEMPM model. In each iteration, the optimization is shown
to be a pseudo-concave problem, which thus guarantees that each
local solution will be the global solution in this step.
The Biased Minimax Probability Machine (BMPM), a global learning
method for biased or imbalanced learning
Present a rigorous and systematic treatment for biased learning tasks
Although being a special case of our proposed general global learning
model, MEMPM, this model provides a quantitative and rigorous
approach for biased learning tasks, where one class of data is always
more important than the other class. Importantly, with explicitly
controlling the accuracy of one class, this branch model can precisely
impose biases on the important class.
Containing explicit generalization bounds for both classes of data

Inheriting the good feature of the MEMPM model, this model also
contains explicit generalization bounds for both classes of data. This
therefore guarantees a good prediction accuracy for future data.
The Local Support Vector Regression (LSVR), a novel regression model
Provide a systematic and automatic treatment in adapting margins
Motivated from M4 , LSVR focuses on considering the margin setting
locally. When compared to the regression model of SVM, i.e. the Support Vector Regression (SVR), this novel regression model is shown
to be more robust with respect to the noise of data in that it contains
the volatile margin setting.
Incorporate special cases very much similar to the standard SVR
When considering a consistent trend for all data points, the LSVR
can derive special cases very much similar to the standard SVR. We
further demonstrate that in a meaningful assumption, the standard
SVR is actually the special case of our LSVR model.
Support Vector Regression with Local Margin Variations
Motivated from the local view of data, another variation of SVR is proposed. It aims to adapt the margin in a more explicit way. This model is
similar to LSVR in the sense that they both adapt margin locally.
We describe the relationship among our developed models in Fig. 1.5.
Fig. 1.5. The relationship among the developed models in this book
1 Introduction
1.5 Scope
This book states and refers to the learning rst as statistical learning, which
appears to be the current main trend of learning approaches. We then further
restrict the learning in the framework of classication, one of the main problems in machine learning. The corresponding discussions on dierent models
including the conducted analysis of the computational and statistical aspects
of machine learning are all subject to the classication tasks. Nevertheless,
we will also extend the content of this book to regression problems, although
it is not the focus of this book.
1.6 Book Organization

The rest of this book is organized as follows:
Chapter 2
We will review dierent learning paradigms in this chapter. We will establish a hierarchy graph attempting to categorize various models in the
framework of local learning and global learning. We will then base this
graph to describe and discuss these models. Finally, we motivate the
Minimum Error Minimax Probability Machine and the Maxi-Min Margin Machine.
Chapter 3
We will develop a novel global learning model, called the Mininum Error
Minimax Probability Machine. We will demonstrate how this new model
represents the worst-case Bayes optimal classier. We will detail its model
denition, provide interpretations, establish a robust version, extend to
nonlinear classications, and present a series of experiments to demonstrate the advantages of this model.
Chapter 4
We will present the Maxi-Min Margin Machine, which successfully combines two dierent but complementary learning paradigms, i.e. local
learning and global learning. We will show how this model incorporates
the Support Vector Machine, the Minimax Probability Machine, and the
Fisher Discriminant Analysis as special cases. We will also demonstrate
the advantages of Maxi-Min Margin Machine by providing theoretical,
geometrical, and empirical investigations.
Chapter 5
An extension of the proposed MEMPM model will be discussed in this
chapter. More specically, the Biased Minimum Minimax Probability Machine will be discussed and applied into the imbalanced learning tasks.
We will review dierent criteria for evaluating imbalanced learning approaches. We will then base these criteria to tailor BMPM into this type
of learning. Both illustrations on toy datasets and evaluations on real
world imbalanced and medical datasets will be provided in this chapter.
References
Chapter 6
A novel regression model called the Local Support Vector Regression,
which can be regarded as an extension from the Maxi-Min Margin Machine, will be introduced in detail in this chapter. We will show that our
model can vary the tube (margin) systematically and automatically according to the local data trend. We will show that this novel regression
model is more robust with respect to the noise of data. Empirical evaluations on both synthetic data and real nancial time series data will
be presented to demonstrate the merits of our model with respect to the
standard Support Vector Regression.
Chapter 7
In this Chapter, we show how to adapt the margin settings locally for
the Support Vector Regression dierently from the LSVR. We demonstrate how the local view of data can be widely used in various models
or even dierently applied in the same model. Empirical evaluations are
also presented in comparison with other competitive models on nancial
data.
Chapter 8
We will then summarize this book and conduct discussions on future
work.
We try to make each of these chapters self-contained. Therefore, in several
chapters, some critical contents, e.g. model denitions or illustrative gures,
having appeared in previous chapters, may be briey reiterated.
References
1. Barber CB, Dobkin DP, Huhanpaa H (1996) The quickhull algorithm for convex
hulls. ACM Transactions on Mathematical Software 22(4):469483
2. Baum LE, Egon JA (1967) An inequality with applications to statistical estimation for probabilistic functions of a Markov process and to a model for
ecology. Bull. Amer. Meteorol. Soc. 73:360C-363
3. Bozdogan H (2004) Statistical Data Mining and Knowledge Discovery. Boca
Raton, Fla.: Chapman & Hall/CRC
4. Christopher J, Burges C (1998) A tutorial on support vector machines for
pattern recognition. Data Mining and Knowledge Discovery 2(2):121167
5. Chow CK, Liu CN (1968) Approximating discrete probability distributions
with dependence trees. IEEE Trans. on Information Theory 14:462467
6. Duda R, Hart P(1973) Pattern Classication and Scene Analysis. New York,
NY: John Wiley & Sons
7. Forsyth DA, Ponce J (2003) Computer Vision: A Modern Approach. Upper
Saddle River, N.J. : Prentice Hall
8. Friedman N, Geiger D, Goldszmidt M (1997) Bayesian network classiers.
Machine Learning 29:131161
9. Fukunaga K (1990) Introduction to Statistical Pattern Recognition. San Diego,
Academic Press, 2nd edition
10
References
10. Girosi F (1998) An equivalence between sparse approximation and support

vector machines. Neural Computation 10(6):14551480
11. Gonzalez MG, Thomason RC (1978) Syntactic Pattern Recognition: An Introduction. Reading, Mass. : Addison-Wesley Pub. Co., Advanced Book Program
12. Grzegorzewski P, Hryniewicz O, Gil M (2002) Soft Methods in Probability,
Statistics and Data Analysis. Heidelberg; New York: Physica-Verlag
13. Hackman D, Meek C, Cooper G (1995) A tutorial on learning bayesian networks. In Tech Report MSR-TR-95-06. Microsoft Research
14. Huang K, King I, Lyu MR (2003) Discriminative training of Bayesian chow-liu
tree multinet classiers. In Proceedings of International Joint Conference on
Neural Network (IJCNN-2003), Oregon, Portland, U.S.A. 1: 484488
15. Jaakkola TS, Haussler D (1998) Exploiting generative models in discriminative
classiers. In Advances in Neural Information Processing Systems (NIPS)
16. Jebara T (2002) Discriminative, Generative and Imitative Learning. PhD
thesis, Massachusetts Institute of Technology
17. Jordan MI (1998) Learning in Graphical Models. Kluwer Academic Publishers
18. Toussaint GT, Jaromczyk JW (1992) Relative neighborhood graphs and their
relatives. Proceedings IEEE 80(9):15021517
19. Lanckriet GRG, Ghaoui LE, Bhattacharyya C, Jordan MI (2002) A robust
minimax approach to classication. Journal of Machine Learning Research
3:555582
20. Marshall AW, Olkin I (1960) Multivariate Chebyshev inequalities. Annals of
Mathematical Statistics 31(4):10011014
21. McLachlan GJ, Basford KE (1988) Mixture Models: Inference and Applications
to Clustering. New York, NY: Marcel Dekker Inc
22. Pearl J (1988) Probabilistic Reasoning in Intelligent Systems: Networks of
Plausible Inference. San Francisco, CA: Morgan Kaufmann
23. Popescu I, Bertsimas D (2001) Optimal inequalities in probability theory: A
convex optimization approach. Technical Report TM62, INSEAD
24. Pruessner A (2003) Conic programming in GAMS. In Optimization Software
The State of the Art. INFORMS Atlanta, http://www.gamsworld.org/cone/links.htm
25. Rabiner LR (1989) A tutorial on hidden Markov models and selected applications in speech recognition. Proceedings of the IEEE 77(2):257-286
26. Russell SJ, Norvig P (1995) Articial Intelligence : A Modern Approach. Englewood Clis, N.J. : Prentice Hall
27. Sch
olkopf B, Smola A (2002) Learning with Kernels. Cambridge, MA: The
MIT Press
28. Smola AJ, Bartlett PL, Scholkopf B, Schuurmans D (2000) Advances in Large
Margin Classiers. MA: The MIT Press
29. Sturm JF(1999) Using sedumi 1.02, a matlab toolbox for optimization over
symmetric cones. Optimization Methods and Software 11:625653
30. Thiesson B, Meek C, Heckman D (1998). Learning mixtures of Bayesian networks. In Technique Report, MSR-TR-97-30. Microsoft Research
31. Vapnik VN (1998). Statistical Learning Theory. John Wiley & Sons
32. Weizenbaum J (1966). Elizaa computer program for the study of natural
language communication between man and machine. Communications of the
Association for Computing Machinery
33. Yedidia J, Freeman WT, Weiss Y (2000). Generalized belief propogation. In
Neural Information Processing Systems 13
References
11
34. Zhang W, King I (2002) A study of the relationship between support vector
machine and Gabriel Graph. In Proceedings of IEEE World Congress on Computational IntelligenceInternational Joint Conference on Neural Networks
2
Global Learning vs. Local Learning
In this chapter, we conduct a more detailed and more formal review on two
dierent schools of learning approaches, namely, the global learning and local
learning. We rst provide a hierarchy graph as illustrated in Fig. 2.1 in which
we try to classify many statistical models into their proper categories, either
global learning or local learning. Our review will also be conducted based on
this hierarchy structure. To make it clear, we use lled shapes to highlight
our own work in the graph.
Global learning ts a distribution over data. If a specic mathematical
model, e.g. a Gaussian model, is assumed on the distribution, this is often
called generative learning, whose name implies that the mathematical formulation of the assumed model governs the generation of data in the learning
task. To learn the parameters from the observations of data for the specic
model, several schemes have been proposed. This includes Maximum Likelihood (ML) learning, which is easy to conduct but is less accurate, Conditional
Likelihood (CL) learning, which is usually hard to perform optimization but
is more eective, and Bayesian Average (BA) learning, which has a comparatively short history but is more promising. As generative learning pre-assigns
a specic model before learning, it often lacks the generality and thus may
be invalid in many cases. This thus motivates the non-parametric learning,
which still estimates a distribution on data but assumes no specic mathematical generative models. The common way in this type of learning is to
locally t over each observation a simple density and then sums all the local
densities as the nal distribution for data. Although in some circumstances,
this approach is successful, it is criticized for requiring a huge quantity of
training points and containing a large space complexity. Dierently, in this
book, we will demonstrate a novel global learning method, named Minimum
Error Minimax Probability Machine (MEMPM). Although still in the framework of global learning, it does not belong to non-parametric learning, therefore requiring no extremely heavy storage spaces. Moreover, it does not
assume any specic distribution on data, which hence distinguishes itself
14
2 Global Learning vs. Local Learning
Fig. 2.1. A hierarchy graph of statistical learning models
from the traditional global generative learning. As a critical contribution,

MEMPM represents a distribution-free Bayes optimal classier in a worstcase scenario. Furthermore, we will show that this model incorporates two
important global learning approaches, Biased Minimax Probability Machine
(BMPM) and Minimax Probability Machine (MPM) [29, 30]. Since all approaches within the paradigm of global learning require summarizing the
data information completely and globally, it thus may waste computational
2.1 Problem Denition
15
resources and is widely argued to be less direct. This motivates the local
learning which makes no attempt to model the data globally, but focuses on
extracting only those information directly related to the task. This type of
learning is often refereed to as discriminative learning in the context of classications. One famous model among them is Support Vector Machine (SVM).
With the task-oriented, robust, computationally tractable properties, SVM
has achieved a great success and is considered as the current state-of-theart classier. Although local learning demonstrates superior performance to
traditional global learning, it appears to situate itself at another extreme
end, which totally discards the useful global information, e.g. the structure
information of data.
Our suggestion is that we should combine these two dierent but complementary paradigms. Towards this end, we then propose a new model called
Maxi-Min Margin Machine (M4 ), which not only successfully employs the
global structure information from data but also holds merits of local learning
such as robustness and superior classication accuracies. As a critical contribution, M4 , the hybrid learning model represents a general model successfully
shown to contain both local learning models and global learning models as
special cases. More specically, it contains two signicant and popular global
learning models, i.e. Fisher Discriminant Analysis (FDA) [13] and Minimax
Probability Machine [28, 29, 30] as special cases. Meanwhile, SVM, the local
learning model can also be considered as one of its branches. In addition,
M4 also demonstrates a strong connection with MEMPM, the novel general
global learning model.
In the following, we rst present the problem denition which will be used
throughout this book. We then base Fig. 2.1 to provide introductions and
comments for each type of learning model sequently. Finally, we summarize
the review and conclude with the proposition of the hybrid framework, the
objective of this book.
2.1 Problem Denition

Given a dataset D consisting of N observations, where each observation is
of the form (z1 , z2 , . . . , zn , c) (zi R, for 1 i n, c F, where F is a
nite set), the basic learning problem is to construct a mapping rule or a
function f from {z1 , z2 , . . . , zn } called features or attributes to the output
c, denoted as the class variable, namely f (z1 , z2 , . . . , zn , , D) c, where
means the function parameters. The function f should be not only as accurate
as possible to t the observations D, but also can robustly predict the class
for the new data. Sometimes, we also use to denote the mapping model
f and its associated parameters. For simplicity, we often use z to denote
the n-dimensional variable {z1 , z2 , . . . , zn }. If we use z j , we refer it to the
j-th observation in D. Throughout this book, unless we provide statements
16
explicitly, and bold typeface will indicate a vector or matrix, while normal
typeface will refer to a scale variable or the component of the vectors.
2.2 Global Learning

Global learning often describes the data by attempting to estimate a distribution over variables (z1 , z2 , . . . , zn , c), denoted as p(z, c, |D). The estimated
distribution can then be used to make predictions by calculating the probability that a specic value of c will occur, when given an instance of features
z. In more details, the decision rule or the mapping function can be described
as:

c = arg max p(ck |D, z) = arg max p(ck , |D, z)d .
(2.1)
ck F
ck F
By employing Bayes theory, one can transform the above joint probability
(the item inside the integral) into the following equivalent forms:
p(ck , z|D, )p(|D)

.
ck F p(ck , z|D, )p(|D)d
p(ck , |D, z) =
(2.2)
Since the denominator in the above does not inuence the decision in
practice, the decision rule of Eq.(2.1) can be written into a relatively easilycalculated form:

c = arg max p(ck , z|D, )p(|D)d .
(2.3)
ck F
Depending on how the model is assumed on D, global learning can

be further divided into generative learning and non-parametric learning as
elaborated in the following subsections.
2.2.1 Generative Learning
Generative learning often assumes a specic model on data D. For example,
a Gaussian distribution is assumed to be the underlying model to generate
D. In this case, the parameters refer to the mean and covariance for the
Gaussian distribution. There are many models which belong to this type of
learning. Among them are Naive Bayes model [9, 26, 32], Gaussian Mixture
Model [4, 15, 16, 33], Bayesian Network [19, 20, 21, 31, 40], Hidden Markov
Model [2, 48], Logistic Regression [23], Bayes Point Machine [18, 36, 44],
Maximum Entropy Estimations [22], etc. The key problem for generative
learning is how to learn the parameters from data. Generally, in the literature of machine learning, three schemes, Maximum Likelihood learning,
Conditional Likelihood learning and Bayesian Average learning, are engaged
for estimating the parameters. We state these approaches one by one in the
following.
2.2 Global Learning
17
2.2.1.1 Maximum Likelihood Learning & Maximum A Posterior

Learning
Considering that it is not always easy to calculate the integral in Eq.(2.3),
earlier researchers often try to compute some approximations of Eq.(2.3)
instead. This motivates the Maximum Likelihood learning and Maximum A
Posterior (MAP) learning [9, 40].
These learning methods replace Eq.(2.3) with the formulation below:
c = arg max p(ck , z|D, ) .
(2.4)
ck F
In the above, how are estimated, thus discriminates MAP from ML.
In MAP, are estimated as:
= arg max p(|D) ,
(2.5)
while in ML, the parameters are given as:

= arg max p(D|) .
(2.6)
Observing Eq.(2.3), one can see that MAP actually enforces the approximated conditional distribution over parameters as a delta function situating
itself at the most prominent . Namely,

1, if = arg max p(|D)
p(|D) =
.
(2.7)
0, otherwise
For ML, it is even simpler. This can be observed by looking into the
relationship between MAP and ML:
arg max p(|D) = arg max p(D|)p() .
(2.8)
Thus, compared to MAP, ML omits the item p(), the prior probability
over the parameters. In practice, a model with a more complex structure
may be more possible to cause over-tting, which means the model can t
the training data perfectly while having a bad prediction ability on the test
or future data. In this sense, discarding the prior probability, ML lacks the
exibility to favor simple models by conditioning the prior probability [5, 49].
On the other hand, MAP permits a regularization on the prior probability
and thus contains potentials to resist over-tting problems.
When applied in practice, under independent, identically distributional
data (i.i.d.) conditions, rather than directly optimizing the original form, ML
estimations usually take the maximization on the log-likelihood, which can
transform the multiplication form into an easily-solved additional one:
= arg max p(D|) = arg max log p(D|) = arg max
N

j=1
log p(z j |). (2.9)
18
2.2.1.2 Maximum Conditional Learning

Rather than computing the integral form, both the above ML learning and
MAP learning seek to use one specic point to calculate Eq.(2.3). The
dierence between them lies in how they estimate the specic parameter
. Compared with the long history in using ML and MAP estimations,
Maximum Conditional (MC) learning enjoys a short span of time but has
achieved state-of-the-art performance in many domains such as speech recognition [4, 42, 53].
Maximum Conditional learning also focuses on adopting one certain
to simplify the computation of Eq.(2.3). Dierently, the selection of is
based on maximizing a conditional likelihood dened as follows:
= arg max p(C|, Z) ,
(2.10)
where C = {c1 , c2 , . . . , cN } is the vector formed by the class label of each

observation in D, and Z = {z 1 , z 2 , . . . , z N } corresponds to the data of the
attributes (or features) in D. Similar to the relation between ML and MAP,
MC can also plug in a prior probability into the above formulae for resisting
over-tting problems, i.e.
= arg max p(C|, Z)p() .
(2.11)
By maximizing the conditional likelihood, MC is thus more direct and

classicationoriented. Note that only the conditional probability which is
maximized above is directly related to the classication purpose. Maximizing
other quantities as done in ML or MAP, possibly optimizes unnecessary information for classications, which is wasteful and imprecise. However, although
MC appears to be more precise, it is usually hard to conduct the optimization due to the involvement of the conditional item. Such an example can be
seen in optimizing a tree-based Bayesian network [12]. Moreover, when there
is missing information, the optimization of MC may even present a more
tough problem in general, while in such circumstances, powerful Expectation
Maximization (EM) techniques [27, 35] can easily be applied in ML.
2.2.1.3 Bayesian Average Learning
It is noted that in ML, MAP and MC, for the easy calculation of Eq.(2.3)
one certain is adopted for approximations. However, although one point
estimation enjoys computational advantages in approximating Eq.(2.3), in
practice it may be very inaccurate and in this sense may impair the prediction
ability of global learning. Aiming to solve this problem, recent researches
have suggested to use the Bayesian Average learning approaches. This type
of approaches facilitates the computation of Eq.(2.3) by changing the integral
into a summation form based on sampling methods, e.g. Markov Chain Monte
Carlo methods [14, 25, 37, 38, 41].
2.2 Global Learning
19
Following this trend, many models are proposed. Among them are Bayesian
Point Machine [18, 36, 44] and Maximum Entropy Estimation [22]. Bayes
Point Machine restricts the averaging of the parameters in the version space
which denotes the space where the training data can be perfectly classied.
This proposed method is reported to contain a better generalization ability
within the global learning framework. But it is challenged to lack systematic
ways to extend its applications into non-separable datasets, where the version
space may include no candidate solutions. Maximum Entropy Estimation, on
the other hand, seems to provide a more exible and more systematic scheme
to perform the averaging of models. By trying to maximize an entropy-like
objective, Maximum Entropy Estimation demonstrates some characteristics
of both global learning and local learning. However, only two small datasets
are used to evaluate its performance. Moreover, the prior, usually unknown,
plays an important role in this model, but has to be assumed beforehand.
2.2.2 Non-parametric Learning
In contrast with generative learning discussed in the above, non-parametric
learning does not assume any specic global models before learning. Therefore, no risk will be taken on possible wrong assumptions on data. Consequently, non-parametric learning appears to set a more valid foundation
than generative learning models. Typical non-parametric learning models in
the context of classications consist of Parzen Window estimation [10] and
the widely used k-Nearest-Neighbor model [7, 43]. We will discuss these two
models in the following.
The Parzen Window estimation also attempts to estimate a density among
the training data. However it employs a totally dierent way. Parzen Window
rst denes an n-dimensional cell hypercube region RN over each observation.
By dening a window function:

1, |uj | 1/2, j = 1, 2, . . . , n
,
(2.12)
w(u) =
0, otherwise
the density is then estimated as:
pN (z) =

N
1 1
z zi
,
w
N i=1 hN
hN
(2.13)
where hN is dened as the length of the edge of RN .

From the above, one can observe that Parzen Window puts a local density over each observation, the nal density is then the statistical result of
averaging all local densities. In practice, the window function can actually
be general functions including the most commonly-used Gaussian function.
Fig. 2.2 illustrates a density estimated by the Parzen Window algorithm.
The k-Nearest-Neighbor method can be cast as designing a special cell
over each observation and then averages all the cell densities as the overall
20
Fig. 2.2. An illustration of Parzen Window estimation
density for data. More specically, the cell volume VN is designed as follows:
let the cell volume be a function of the training data, by centering a cell
around each point z j and increasing the volume until kN samples are contained, where kN depends on N . The local density for each observation is
then dened as
pN (z j ) =
kN /N
.
VN
(2.14)
When used for classications, the prediction is given by the class with the
maximum posterior probability, i.e.
c = arg max pN (ci |z) .
ci F
(2.15)
Further, the posterior probability can be calculated as below:

pN (ci , z)
(ki /N )/V
ki
=
=
.
pN (ci |z) =
pN (z, ci )
(ki /N )/V
k
iF
(2.16)
iF
Therefore, the prediction result is just the class with the maximum fraction
of the samples in a cell.
These non-parametric methods make no underlying assumptions on data
and appear to be more general in real cases. However, using no parameters
actually means using many parameters so that each parameter would not
dominate other parameters (in the discussed models, the data points can
be in fact considered as the parameters). In such a way, if one parameter
fails to work, it will not inuence the whole system globally and statistically.
However, using many parameters also results in serious problems. One of
the main problems is that the density is overwhelmingly dependent on the
training samples. Therefore, to generate an accurate density, the number of
samples needs to be very large (much larger than would be required if we perform the estimation by generative learning approaches). What is even worse
2.2 Global Learning
21
is that the number of data will unfortunately increase exponentially with the
dimension of data. Another disadvantage caused is its severe requirement for
the storage, since all the samples need to be saved beforehand in order to
predict new data.
2.2.3 The Minimum Error Minimax Probability Machine
Within the context of global learning, a dilemma seems existing: If we assume
a specic model as in generative learning, it loses the generality; if we use
instead non-parametric learning, it is impractical for high-dimension data.
One question is then proposed, can we have an approach which does not
require a large number of training samples for reducing complexities and also
does not assume specic models for maintaining the generality? Towards this
end, we propose Minimum Error Minimax Probability Machine (MEMPM)
in this book.
Unlike generative learning or non-parametric learning, Minimum Error
Minimax Probability Machine does not try to estimate a distribution over
data. Instead, it attempts to extract reliable global information from data and
estimates parameters for maximizing the minimal possibility that a future
data will fall into the correct class. More precisely, rather than seeking to
nd an accurate distribution, MEMPM focuses on studying the worst-case
probability (which is relatively robust) to predict data. In terms of the style
in making decisions, MEMPM is more like a local learning method due to
its direct optimization for classication and the task-oriented characteristic.
However, because MEMPM only summarizes global information from data
(not a distribution) as well, we still locate it in the framework of global
learning.
The proposed MEMPM contains many appealing features. Firstly, it represents a distribution-free Bayes optimal classier in the worst-case scenario.
A perfect balance is achieved by MEMPM in this way: No specic model is
assumed on data, since it is distribution-free. At the same time, although in
the worst-case scenario, it is also the Bayes optimal classier which is only
originally applicable in the cases with a known distribution. Another critical
feature of MEMPM is that under a mild condition, it contains an explicit
generalization bound. Furthermore, by exploring the bound, the recentlyproposed promising model, Minimax Probability Machine is clearly demonstrated to be its special case. Importantly, based on specifying a bound for
one class of data, a Biased Minimax Probability Machine is branched out
from MEMPM, which will be shown to provide a rigorous and systematic
treatment for biased classications. We will detail the MEMPM model and
BMPM model in the next chapter.
22
2.3 Local Learning

Local learning adopts a largely dierent way to construct classiers. This
type of learning is even more task-oriented than Minimum Error Minimax
Probability Machine and Maximal Conditional learning. In the context of
classications, only the nal mapping function from the features z to c is
crucial. Therefore, describing global information from data or explicitly summarizing a distribution whatever is conditional or joint, is a roundabout or
intermediate step and therefore may be deemed wasteful or imprecise especially when the global information cannot be estimated accurately.
Alternatively, recent progress has suggested a local learning method, or
well known as the discriminative learning method. The family of approaches
directly pin-points the most critical quantities for classications, while all
other information less irrelevant to this purpose is simply omitted. Compared
to global learning, no model is assumed and also no explicit global information
will be engaged in this scheme. Among this school of methods are Neural
Networks [1, 11, 17, 34, 39, 43], Gabriel Graph methods [3, 24, 54], large
margin classiers [8, 45, 46, 47] including Support Vector Machine (SVM),
a state-of-the-art classier which achieves superior performance in various
pattern recognition elds. In the following, we will focus on introducing SVM
in details.
Support V ector M achines
Support Vector Machine is established based on minimizing the expected
classication risk as dened as follows:

R() =
l(z, c, )d(p(z, c)) ,
(2.17)
z,c
where l(z, c, ) is the loss function. Similar problems occur in the global
learning, since generally p(z, c) is unknown. Therefore, in practice, the above
expected risk is often approximated by the so-called empirical risk:
Remp () =
N
1 j j
l(z , c , ) .
N j=1
(2.18)
The above loss function describes the extent on how close the estimated
class disagrees with the real class for the training data. Various metrics can be
used for dening this loss function, including the 0 1 loss and the quadratic
loss [50].
However, considering only the training data may lead to the over-tting
problem again. In SVM, one big step in dealing with the over-tting problem
has been made, i.e. the margin between two classes should be pulled away
in order to reduce the over-tting risk. Fig. 2.3 illustrates the idea of SVM.
2.4 Hybrid Learning
23
Fig. 2.3. An illustration of Support Vector Machine
Two classes of data depicted as circles and solid dots are presented in this
gure. Intuitively observed, there are many decision hyperplanes which can be
adopted for separating these two classes of data. However, the one plotted in
this gure is selected as the favorable separating plane, because it contains the
maximum margin between two classes. Therefore, in the objective function
of SVM, a regularization term representing the margin shows up. Moreover,
as seen in this gure, only those lled points called support vectors mainly
determine the separating plane, while other points do not contribute to the
margin at all. In another word, only several local points are critical for the
classication purpose in the framework of SVM and thus should be extracted.
Actually, a more formal explanation and theoretical foundation can be
obtained from the Structure Risk Minimization criterion [6, 52]. Therein,
maximizing the margin between dierent classes of data is minimizing an
upper bound of the expected risk, i.e. the VC dimension bound [52]. However,
since the focus of this book does not lie in the theory of SVM, we will not go
further to discuss the details about this. Interested readers can refer to [51,
52].
2.4 Hybrid Learning

Local learning (or simply regarded as SVM) has demonstrated its advantages,
such as its state-of-the-art performance (the lower generalization error), the
optimal and unique solution, and the mathematical tractability. However, it
does discard many useful information from data, e.g. the structure information from data.
An illustrative example has been seen in Fig. 1.4. In the current stateof-the-art classier, i.e. SVM, similar problems also occur. This can be seen
in Fig. 2.4. In this gure, the purpose is to separate two catergories of data
x and y. As observed, the classication boundary is intuitively observed to
be mainly determined by the dotted axis, i.e. the long axis of the y data
24
(represented by s) or the short axis of the x data (represented by s).

Moreover, along this axis, the y data are more possible to scatter than the x
data, since y contains a relatively larger variance in this direction. Noting this
global fact, a good decision hyperplane seems reasonable to lie closer to the
x side (see the dash-dot line). However, SVM ignores this kind of global
information, i.e. the statistical trend of data occurrence. The derived SVM
decision hyperplane (the solid line) lies unbiasedly right in the middle of
two local points (the support vectors).The above considerations directly
motivate Maxi-Min Margin Machine.
Fig. 2.4. A decision hyperplane with considerations of

both local and global information
2.5 Maxi-Min Margin Machine

After examining the road-map of the learning models, especially the global
learning and local learning, we have seen a strong motivation for combining
two dierent but complementary schemes. More specically, borrowing the
idea from local learning by assuming no distribution on data would set a
valid foundation for the learning models. Meanwhile, fusing robust global
information, e.g. structure information, into learning models appears to benet more on rening decisions in separating data.
Our eort will be made in this direction. As will be detailed in Chapter 4, the hybrid learning model, Maxi-Min Margin Machine successfully plugs
the global information into the learning and enjoys good features from both
local learning and global learning. As seen in Fig. 2.1, the Maxi-Min Margin Machine model has built up various connections with many models in
the literature; it incorporates Support Vector Machine as a special case,
which lies in the framework of local learning; it also includes Minimax
References
25
Probability Machine and Fisher Discriminant Analysis as direct spin-os.

Moreover, a strong link has been established between this model and Minimum Error Minimax Probability Machine. Moreover, empirical investigations
have shown that this combined model outperforms both local learning model
such as SVM and global learning models, e.g. MPM.
In the next chapter, we will rst present the Minimum Error Minimax
Probability Machine which is a general global learning model. Following that,
we then introduce the Maxi-Min Margin Machine and demonstrate its merits
both theoretically and empirically.
References
1. Anand R, Mehrotram GK, Mohan KC, Ranka S (1993) An improved alogrithm
for neural network classication of imbalance training sets. IEEE Transactions
on Neural Networks 4(6):962969
2. Bahl LR, Brown PF, de Souza PV, Mercer RL (1993) Estimating hidden
Markov model parameters so as to maximize speech recognition accuracy. IEEE
Transactions on Speech and Audio Processing 1:7782
3. Barber CB, Dobkin DP, Huhanpaa H (1996) The quickhull algorithm for convex
hulls. ACM Transactions on Mathematical Software 22(4):469483
4. Beaufays F, Wintraub M, Konig Y (1999) Discriminative mixture weight estimation for large Gaussian mixture models. In Proceedings of the International
Conference on Acoustics, Speech and Signal Processing 337340
5. Brand M (1998) Structure discovery via entropy minimization. In Neural
Information Processing System 11
6. J Christopher, Burges C (1998) A tutorial on support vector machines for
pattern recognition. Data Mining and Knowledge Discovery 2(2):121167
7. Cover TM, Hart PE (1967) Nearest neighbor pattern classication. IEEE
Transactions on Information Theory IT-13(1):2127
8. Cristianini N, Shawe-Taylor J (2000) An Introduction to Support Vector Machines(and Other Kernel-based Learning Methods). Cambridge, U.K.; New
York, NY: Cambridge University Press
9. Duda R, Hart P (1973) Pattern Classication and Scene Analysis. New York,
NY: John Wiley & Sons
10. Duda RO, Hart PE, Stork DG (2000) Pattern Classication. New York, NY:
John Wiley & Sons
11. Fausett L (1994) Fundamentals of Neural Networks. New York, NY: Prentice
Hall
13. Fukunaga K (1990) Introduction to Statistical Pattern Recognition. San Diego,
Academic Press, 2nd edition
14. Gilks WR, Richardson S, Spiegelhalter DJ (1996) Markov Chain Monte Carlo
in Practice. London: Chapman & Hall
15. Grzegorzewski P, Hryniewicz O, Gil M (2002) Soft Methods in Probability,
Statistics and Data Analysis. Heidelberg; New York: Physica-Verlag
26
References
16. Hastie T, Tibshirani R (1996) Discriminant analysis by Gaussian mixtures.

Journal of the Royal Statistical Society(B) 58:155176
17. Haykin S (1994) Neural Networks: A Comprehensive Foundation. New York,
NY: Macmillan Publishing
18. Herbrich R, Graepel T (2001) Large scale Bayes point machines. In Advances
in Neural Information Processing Systems (NIPS)
19. Huang K, King I, Chan L, Yang H (2004) Improving Chow-Liu tree performance
based on association rules. In J. C. Rajapakse and L. Wang, editors, Neural
Information Processing: Research and Development, Studies in Fuzziness and
Soft Computing, 152: 94112. Heidelberg; New York: Springer-Verlag
20. Huang K, King I, Lyu MR (2002). Learning maximum likelihood semi-naive
Bayesian network classier. In Proceedings of IEEE International Conference
on Systems, Man and Cybernetics (SMC2002). Hammamet, Tunisia TA1F3
21. Huang K, King I, Lyu MR (2003) Finite mixture model of bound semi-naive
Bayesian network classier. In Proceedings of the International Conference on
Articial Neural Networks (ICANN-2003), Lecture Notes in Articial Intelligence, Long Paper. Heidelberg: Springer-Verlag 2714: 115122
22. Jebara T (2002) Discriminative, Generative and Imitative Learning. PhD
thesis, Massachusetts Institute of Technology
23. Jordan MI (1995) Why the logistic function? A tutorial discussion on probabilities and neural networks. Technical Report 9503, MIT Computational
Cognitive Science Report
24. Toussaint GT, Jaromczyk JW(1992) Relative neighborhood graphs and their
relatives. Proceedings IEEE 80(9):15021517
25. Kass RE, Carlin BP, Gelman A, Neal RM (1998) Markov chain Monte Carlo
in practice: A roundtable discussion. The American Statistician 52:93100
26. Kohavi R, Becker B, Sommereld D (1997) Improving simple Bayes. In Technique Report. Mountain View, CA: Data Mining and Visualization Group,
Silicon Graphics Inc
27. Laird NM, Dempster AP, Rubin DB (1977) Maximum likelihood from incomplete data via the EM algorithm. J. Royal Statist. Society B39:138
28. Lanckriet GRG, Ghaoui LE, Bhattacharyya C, Jordan MI (2001) Minimax
probability machine. In Advances in Neural Information Processing Systems
(NIPS)
3:555582
30. Lanckriet GRG, Ghaoui LE, Jordan MI (2002) Robust novelty detection with
single-class MPM. In Advances in Neural Information Processing Systems
(NIPS)
31. Langley P (1993) Introduction of recursive Bayesian classiers. In Proceedings
of the 1993 European Conference on Machine Learning 153164
32. Langley P, Iba W, Thompson K (1992) An analysis of Bayesian classiers. In
Proceedings of National Conference on Articial Intelligence 223228
33. McLachlan GJ, Basford KE (1988) Mixture Models: Inference and Applications
to Clustering. New York, NY: Marcel Dekker Inc
34. Pankaj Mehra, Benjamin W Wah (1992) Articial Neural Networks : Concepts
and Theory. Los Alamitos, California: IEEE Computer Society Press
35. Meng XL, Rubin DB (1993) Maximum likelihood estimation via the ECM
algorithm: A general framework. Biometrika 80(2)
References
27
36. Minka T (2001) A family of Algorithms for Approximate Inference. PhD thesis,
Massachusetts Institute of Technology
37. Neal RM (1993) Probabilistic inference using Markov chain Monte Carlo methods. Technical Report CRG-TR-93-1, Dept. of Computer Science, University
of Toronto
38. Neal RM (1998). Suppressing random walks in Markov chain Monte Carlo using
ordered overrelaxation M. I. Jordan (editor) Learning in Graphical Models,
Dordrecht: Kluwer Academic Publishers 205225
39. Patterson D (1996) Articial Neural Networks. Singapore: Prentice Hall
40. Pearl J (1988) Probabilistic Reasoning in Intelligent Systems: Networks of
Plausible Inference. San Francisco, CA: Morgan Kaufmann
41. Pinto RL, Neal RM (2001) Improving Markov chain Monte Carlo estimators
by coupling to an approximating chain. Technical Report No. 0101, Dept. of
Statistics, University of Toronto
42. Rathinavelu C, Deng L (1996) The trended HMM with discriminative training
for phonetic classication. In Proceedings of ICSLP
43. Ripley BD (1996) Pattern Recognition and Neural Networks. Press Syndicate
of the University of Cambridge
44. Rujam R (1997) Preceptron learning by playing billiards. Neural Computation
9:99122
45. Sch
olkopf B, Burges C, Smola A (1999) Advances in Kernel Methods: Support
Vector Learning. Cambridge, MA: The MIT Press
46. Sch
olkopf B , Smola A (2002) Learning with Kernels: Support Vector Machines,
Regularization, Optimization and Beyond. Cambridge, MA: The MIT Press
47. Smola AJ, Bartlett PL, Scholkopf B, Schuurmans D (2000). Advances in Large
Margin Classiers. Cambridge, MA: The MIT Press
48. Stolcke A, Omohundro S (1993) Hidden Markov model induction by Bayesian
model merging. In NIPS 5: 1118
49. Tipping M(1999) The relevance vector machine. In Advances in Neural Information Processing Systems 12 (NIPS)
50. Trivedi PK (1978) Estimation of a distributed lag model under quadratic loss.
Econometrica 46(5):11811192
51. Vapnik VN (1998) Statistical Learning Theory. New York, NY: John Wiley &
Sons
52. Vapnik VN (1999) The Nature of Statistical Learning Theory. New York, NY:
Springer, 2nd edition
53. Woodland P, Povey D (2000) Large scale discriminative training for speech
recognition. In Proceedings of ASR 2000
54. Zhang W, King I (2002) A study of the relationship between support vector
machine and Gabriel Graph. In Proceedings of IEEE World Congress on Computational IntelligenceInternational Joint Conference on Neural Networks
3
A General Global Learning Model: MEMPM
Traditional global learning, especially generative learning, enjoys a long and

distinguished history, holding a lot of merits, e.g. a relatively simple optimization, and the exibility in incorporating global information such as structure information and invariance, etc. However, it is widely argued that this
model lacks the generality for having to assume a specic model beforehand.
Assuming a specic model over data is useful in some cases. However, the assumption may not always coincide with the true data distribution in general
and thus may be invalid in many circumstances. In this chapter, we propose
a novel global learning model, named Minimum Error Minimax Probability
Machine (MEMPM), which is directly motivated from Marshall and OlKin
Probability Theory [20, 24]. For classifying data correctly, this model focuses
on estimating the worse-case probability, which is not only more reliable,
but also more importantly provides no need for assuming specic models.
Furthermore, this new model consists of several appealing features.
First, MEMPM acutally presents a novel general framework for classications. As demonstrated later, MEMPM includes a recently-proposed promising model Minimax Probability Machine as its special case, which is reported
to achieve comparable performance to SVM. Interpretations from both viewpoints of the optimal thresholding problem and the geometry will be provided
to show the advantages of MEMPM. Moreover, this novel model branches out
another promising special case, named Biased Minimax Probability Machine
(BMPM) [12] and extends its application into a type of important classications, i.e. biased classications.
Second, this model derives a distribution-free Bayes optimal classier
in the worst-case scenario. It thus distinguishes itself from the traditional
global learning methods, or more particularly, the traditional Bayes optimal
classiers which have to assume a distribution on data and thus lack the
generality in real cases. Furthermore, we will show that under some conditions, e.g. when a Gaussian distribution is assumed on data, the worst-case
Bayes optimal classier becomes the true Bayes optimal hyperplane.
30
3 A General Global Learning Model: MEMPM
Third, the MEMPM model contains an explicit performance indicator,

namely an explicit upper bound on the probability of misclassication of
future data. Moreover, we will demonstrate theoretically and empirically that
MEMPM attains a smaller upper bound of the probability of misclassication
than MPM, which thus implies the advantages of MEMPM over MPM.
Fourth, although in general the optimization of MEMPM is shown to
be a non-concave problem, empirically, it demonstrates a good concavity in
the main interest region and thus can be solved practically. Furthermore,
we will show that the nal optimization problem involves solving a onedimensional line search problem and thus results in a satisfactory solving
method.
This chapter is organized as follows. In the next section, we will rst introduce the Marshall and Olkin Theory. We then present the main content
of this chapter, the MEMPM model, including its denition, interpretations,
the practical solving method, and the sucient conditions for the convergence into the true Bayes decision hyperplane. Following that, we demonstrate a robust version of MEMPM. In Section 3.4, we seek to kernelize the
MEMPM model to attack nonlinear classication problems. We then, in Section 3.5, present a series of experiments on synthetic datasets and real-world
benchmark data sets. In Section 3.6, we analyze the tightness of the worstcase accuracy bound. In Section 3.7, we show that empirically MEMPM is
often concave in the main interest region. In Section 3.8, we present the
limitations of MEMPM and envision the possible future work. Finally, we
summarize this chapter in Section 3.9.
3.1 Marshall and Olkin Theory

The Marshall and Olkin Theory can be described as follows:
Theorem 3.1. [Marshall and Olkin Theory] The probability that a random
vector y belongs to a convex set S can be bounded by the following formulation:
sup
y(y, y )
P r{y S} =
1
, with d2 = inf (y y)T 1
y (y y) , (3.1)
yS
1 + d2
where the supremum is taken over all distributions for y containing the mean
as y and the covariance matrix as y 1 .
The theory provides us with a possibility to assume no model, but bound
the probability of misclassifying a point and consequently develop a novel
classier within the framework of global learning. More specically, one can
design a linear separating plane by replacing S with a half space associated
1
We assume y to be positive denite for simplicity. Otherwise, we can always
add a small positive amount to its diagonal elements to force its positive denition.
3.2 Minimum Error Minimax Probability Decision Hyperplane
31
with this linear plane. To take the supremum can then be considered to
bound the misclassication rate for one class of data. We in the following,
rst introduce the model denition and then show how this theory can be
applied therein for deriving a distribution-free classier.
3.2 Minimum Error Minimax Probability Decision

Hyperplane
In this section, we rst present the model denition of MEMPM while reviewing the original MPM model. We then in Section 3.2.2 interpret MEMPM
with respect to MPM. In Section 3.2.3, we specialize the MEMPM model
for dealing with biased classications. In Section 3.2.4, we analyze the
MEMPM optimization problem and propose a practical solving method. In
Section 3.2.5, we address the sucient conditions when the worst-case Bayes
optimal classier derived from MEMPM becomes the true Bayes optimal classier. In Section 3.2.6, we provide a geometrical interpretation for BMPM and
MEMPM.
3.2.1 Problem Denition
The notation in this chapter will largely follow that of [16]. Let x and y
denote two random vectors representing two classes of data with means and
covariance matrices as {x, x } and {y, y }, respectively, in a two-category
classication task, where x, y, x, y Rn , and x , y Rnn .
Assuming {x, x }, {y, y } for two classes of data are reliable, MPM
attempts to determine the hyperplane wT z = b (w Rn \{0}, z Rn ,
b R, and superscript T denotes the transpose) which can separate two
classes of data with the maximal probability. The formulation for the MPM
model is written as follows:
max { + (1 )} ,
,,w=0,b
s.t.
(3.2)
inf
P r{wT x b} ,
(3.3)
inf
P r{wT y b} ,
(3.4)
x(x, x )
y(y, y )
where and indicate the worst-case classication accuracies of future data

points for the class x and y, respectively, namely, the worst-case accuracy for
classifying x data and y data. Future points z for which wT z b are then
classied as the class x; otherwise they are judged as the class y. [0, 1] is
the prior probability of the class x and 1 is thus the prior probability of
the class y. Intuitively, maximizing + (1 ) can be naturally considered
as maximizing the expected worst-case accuracy for future data. In other
words, this optimization leads to minimizing the expected upper bound of
32
the error rate. More precisely, if we change max{ + (1 )} to min{(1

)+(1)(1)} and consider 1 as the upper bound probability that an
x data is classied into class y (1 is similarly considered), the MEMPM
model exactly minimizes the maximum Bayes error and thus derives the
Bayes optimal hyperplane in the worst-case scenario. In comparison, MPM
assumes the equal worst-case probability for both classes, i.e. it forces = .
Obvisouly, this is inappropriate since it is unnecessary that the worst-case
accuracies are presumed equal. However, even in such a constrained way,
MPM is reported to achieve comparable performacne to SVM, a current
state-of-the-art classier. Therefore, the generalized case of MPM, namely,
MEMPM may be expected to be more pomising. This will be empirically
demonstrated in the experimental part of this chapter.
3.2.2 Interpretation
We interpret MEMPM with respect to MPM in this section. First, it is evident
that if we presume = , the optimization of MEMPM degrades to the
MPM optimization. This would mean that MPM is actually a special case of
MEMPM.
An analogy to illustrate the dierence between MEMPM and MPM can
be seen in the optimal thresholding problem. Fig. 3.1 illustrates this analogy.
To separate two classes of one-dimensional data with density functions as p1
and p2 , respectively, the optimal thresholding is given by the decision plane
in Fig. 3.1(a) (assuming that the prior probabilities for two classes of data
are equal). This optimal thesholding corresponds to the point minimizing the
(a)
(b)
Fig. 3.1. An analogy to illustrate the dierence between MEMPM

and MPM with equal prior probabilities for two classes. The optimal
decision plane corresponds to the intersection point, where the error
(1 ) + (1 ) is minimized (or the accuracy + is maximized)
as implied by MEMPM, rather than the one where is equal to as
implied by MPM
33
error rate (1 ) + (1 ) or maximizing the accuracy + , which is exactly

the intersection point of two density functions (1 represents the area of
135 -line lled region and 1 represents the area of 45 -line lled region).
On the other hand, the thresholding point to force = is not necessarily
the optimal point to separate these two classes.
It should be claried that the MEMPM model assumes no distributions.
This distinguishes the MEMPM model from the traditional Bayes optimal
thresholding method which has to make specic assumptions on data distribution. On the other hand, although MEMPM minimizes the upper bound
of the Bayes error rate of future data points, as shown later in Section 3.2.5,
it will represent the true Bayes optimal hyperplane under some conditions,
e.g. when a Gaussian distribution is assumed on data.2
3.2.3 Special Case for Biased Classications
The above discussion only covers the unbiased classication tasks, which does
not favor one class over the other class intentionally. However, another important type of pattern recognition tasks, namely biased classication, arises
very often in practice. In this scenario, one class is usually more important
than the other class. Thus a bias should be imposed towards the important
class. Such typical example can be seen in the diagnosis of epidemical disease. Classifying a patient who is infected with a disease into an opposite
class results in serious consequence. Thus in this problem, the classication
accuracy should be biased towards the class with disease. In other words, we
would prefer to diagnose the person without the disease to be the infected
case rather than the other way round.
We in the following demonstrate that MEMPM actually contains a special
case we call Biased Minimax Probability Machine for biased classications.
We formulate this special case as:
max
,,w=0,b
s.t.
inf
P r{wT x b} ,
inf
P r{wT y b} 0 ,
x(x, x )
y(y, y )
2
Another interpretation of the dierence between MEMPM and MPM can be
stated from the viewpoint of Game Theory. MPM can be regarded as a noncooperative competitive game. In this game, each player (class) tries to maximize
its individual benet, i.e. . The competition leads to each class obtaining the same
benet when all classes fulll a kind of equilibrium. However, in the game theory,
many models, e.g. the prisoners dilemma, Counot Model and the tragedy of the
commons [21], have stated that maximizing individual benet does not lead to
maximizing the global optimum. Our model, on the contrary, can be considered as
a kind of cooperative game. It achieves the global optimum through cooperation.
34
where 0 is a pre-specied positive constant, which represents an acceptable

accuracy level for the less important class y.
The above optimization utilizes a typical setting in biased classications,
i.e. the accuracy for the important class (associated with x) should be as high
as possible, if only the accuracy for the less important class (associated with
y) maintains at an acceptable level specied by the lower bound 0 (which
can be set by users).
With quantitatively plugging a specied bias 0 into classications and
also containing an explicit accuracy bound for the important class, BMPM
provides a more direct and elegant way for biased classications. Comparatively, to achieve a specied bias, traditional biased classiers such as the
Weighted Support Vector Machine [23] and the Weighted k-Nearest Neighbor
method [18] usually adapt dierent costs for dierent classes. However, due
to the diculties in building up quantitative connections between the cost
and the accuracy,3 for imposing a specied bias, these methods need resort
to the trial and error procedure to attain suitable costs which are generally
indirect and lack rigorous treatments.
3.2.4 Solving the MEMPM Optimization Problem
In this section, we will propose to solve the MEMPM optimization problem. As will be demonstrated shortly, the MEMPM optimization can be
transformed into a one-dimensional line search problem. More specically,
the objective function of the line search problem is implicitly determined by
dealing with a BMPM problem. Therefore, solving the line search problem
corresponds to solving a Sequential Biased Minimax Probability Machine
(SBMPM) problem. Before we proceed, we rst introduce how to solve the
BMPM optimization problem.
3.2.4.1 Solving the BMPM Optimization Problem
First, we describe Lemma 3.2 which is developed in [16].
Lemma 3.2. Given w = 0 and b, such that wT y b and [0, 1), the
condition:
inf
y(y, y )
P r{wT y b}
.
holds if and only if b wT y () wT y w with () = 1
The lemma can be proved according to the Marshall and Olkin Theory
and the Lagrangian Multiplier theory.
3
Although cross validations could be used to provide empirical connections, they
are problem-dependent and are usually slow procedures as well.
35
Proof. In Marshall and Olkin Theory, if we dene S = {wT y b}, the

theorem is changed to:
sup
y{y, y }
Pr {wT y b} =
1
,
1 + d2
with d2 =
inf (y y)T 1
y (y y) .
wT yb
We next show that d can be obtained as follows:

d2 =
inf (y y)T 1
y (y y) =
wT yb
max (b wT y, 0)2
.
wT y w
This can be proved by using the Lagrangian multiplier method as follows:

(1) If wT y b.
Denoting pT = wT y1/2 , g = y1/2 (y y), and q = b wT y, one
can write d2 = inf g T g. One can obtain g by introducing Lagrangian
pT wq
multiplier:
{g, } = arg min arg max{g T g + (q pT g)},
g
where the multiplier 0. Therefore, one can get the following equalities:
g=
p
,
2
q = pT g.
Since wT y b, one can easily obtain q 0. One can further obtain:

=
2q
,
pT p
g=
dp
.
pT p
Finally, this leads to the following equation:

d2 =
inf (y y)T 1
y (y y) =
wT yb
(b wT y)2
.
wT y w
(2) If wT y b.
In this case, we can only have y = y. Therefore, d = 0.
By integrating the above, we thus complete the proof of this theorem.
By using Lemma 3.2 we can transform the BMPM optimization problem
as follows:
max
,w=0,b
s.t.
b + wT x () wT x w ,

b wT y (0 ) wT y w ,
(3.5)
(3.6)
(3.7)
36
where () =
1 ,

(0 ) =
0
10 .
Eq.(3.7) is directly obtained from
Eq.(3.4) by using Lemma 3.2. Similarly, by changing wT x b to wT (x)

b, Eq.(3.6) can be obtained from Eq.(3.3).
From Eqs.(3.6) and (3.7), we get

wT y + (0 ) wT y w b wT x () wT x w .
(3.8)
If we eliminate b from this inequality, we obtain:

wT (x y) () wT x w + (0 ) wT y w .
(3.9)
We observe that the magnitude of w does not inuence the solution of

Eq.(3.9). Moreover, we can assume x = y; otherwise, if x = y, the minimax machine does not have a physical meaning. In this case, Eq.(3.9) may
even have no solution for every 0 = 0, since the right hand side would be
always positive provided that w = 0. Thus in the extreme case, and have
to be zero, which means the worst-case misclassication are always zero.
Without loss of generality, we can set wT (x y) = 1. Thus the problem
can be further changed as:
max
,w=0
s.t.
(3.10)
1 () wT x w + (0 ) wT y w ,
(3.11)
w (x y) = 1 .
(3.12)
Since x can be assumed as positive denite (otherwise, we can always

add a small positive amount to its diagonal elements and make it positive
denite), from Eq.(3.11) we can obtain:
1 (0 ) wT y w
()
.
(3.13)
wT x w
Because () increases monotonically with , maximizing is equivalent
to maximizing (), which further leads to:
1 (0 ) wT y w
,
max
w=0
wT x w
s.t.
wT (x y) = 1 .
This kind of optimization is called Fractional Programming (FP) problem [13, 19, 26]. To elaborate further, this optimization is equivalent to solving
the following fractional problem:
max
w=0
f (w)
,
g(w)
(3.14)
37
T
T
subject to
w A = {w|w (xy) = 1}, where f (w) = 1(0 ) w y w,

g(w) = wT x w.
Theorem 3.3. The Fractional Programming problem Eq.(3.14) associated
with the BMPM optimization is a pseudo-concave problem whose every local optimum is the global optimum.
Proof. It is easy to see that the domain A is a convex set on Rn , f (w)
and g(w) are dierentiable on A. Moreover, since x and y can be both
considered as positive denite matrices, f (w) is a concave function on A and
g(w) is a convex function on A. Then f (w)/g(w) is a concave-convex FP
problem. Hence it is a pseudo-concave problem [26]. Therefore, every local
maximum is the global maximum [26].
To handle this specic FP problem, many methods such as the parametric
method [26], the dual FP method [7, 25], and the concave FP method [6] can
be used. A typical Conjugate Gradient method [2] in solving this problem will
have a worst-case O(n3 ) time complexity. Adding the time cost to estimate
x, y, x , and y , the total cost for this method is O(n3 + N n2 ), where N is
the number of data points. This complexity is in the same order as the linear
Support Vector Machines [27] and the linear MPM [16].
In this chapter, the Rosen gradient projection method [2] is used to nd
the solution of this pseudo-concave FP problem, which is proved to converge
to a local maximum with a worse-case linear convergence rate. Moreover, the
local maximum will exactly be the global maximum in this problem.
3.2.4.2 Sequential BMPM Optimization Method for MEMPM
We now turn to solving the MEMPM problem. Similar to Section 3.2.4.1, we
can base on Lemma 3.2 to transform the MEMPM optimization as follows:
{ + (1 )} ,
b + wT x () wT x w ,

b wT y () wT y w .
max
,,w=0,b
s.t.
(3.15)
(3.16)
(3.17)
Using the similar analysis as in Section 3.2.4.1, we can further transform

the above optimization into
max { + (1 )} ,
,,w=0
s.t.
(3.18)
1 () wT x w + () wT y w ,
(3.19)
wT (x y) = 1 .
(3.20)
In the following we provide a lemma to show that the MEMPM solution

is actually attained on the boundary of the set formed by the constraints of
Eqs.(3.19) and (3.20).
38
Lemma 3.4. The maximum value of + (1 ) under the constraints of

Eqs.(3.19) and (3.20) is achieved when the right hand side of Eq.(3.19) is
strictly equal to 1.
Proof. Assume the maximum is achieved when

1 > () wT y w + () wT x w .
A new solution constructed by increasing or () by a small positive
amount,4 and maintaining , w unchanged will satisfy the constraints and
will be a better solution.
By applying Lemma 3.4 we can transform the optimization problem
Eq.(3.18) under the constraints of Eqs.(3.19) and (3.20) as follows:

2 ()
max
+ (1 ) ,
(3.21)
,w=0
2 () + 1
s.t.
where
wT (x y) = 1 ,
(3.22)
1 () wT y w
() =
wT x w
.
In Eq.(3.22), if we x to a specic value within [0, 1), the optimization
is equivalent to maximizing 2 ()/2 () + 1 and further equivalent to maximizing (), which is exactly the BMPM problem. We can then update
according to some rules and repeat the whole process until an optimal is
found. This is also the so-called line search problem [2, 1]. More precisely,
if we denote the value of optimization as a function f (), the above procedure corresponds to nding an optimal to maximize f (). Instead of using
an explicit function as in traditional line search problems, the value of the
function here is implicitly given by a BMPM optimization procedure.
Many methods can be used to solve the line search problem. In this
chapter, we use the Quadratic Interpolation (QI) method [2]. As illustrated
in Fig.3.2, QI nds the maximum point by updating a three-point pattern
(1 , 2 , 3 ) repeatedly. The new denoted by new is given by the quadratic
interpolation from the three-point pattern. Then a new three-point pattern
is constructed by new and two of 1 , 2 , 3 . This method can be shown to
converge superlinearly to a local optimum point [2]. Moreover, as shown in
Section 3.7, although MEMPM generally cannot guarantee its concavity, empirically it is often a concave problem. Thus the local optimum will be often
the global optimum in practice.
4
Since () increases monotonically with , increasing by a small positive
amount corresponds to increasing () by a small positive amount.
39
Fig. 3.2. A three-point pattern and quadratic line search method.

A new is obtained and a new three-point pattern is constructed
by new and two of 1 , 2 and 3
Until now, we do not mention how to calculate the intercept b. From

Lemma 3.4, we can see that the inequalities Eqs.(3.16) and (3.17) will become
equalities at the maximum point (w , b ). The optimal b will thus be obtained
by

T w = w T y + ( ) w T w .
b = w T
x
(
)
w
(3.23)
3.2.5 When the Worst-case Bayes Optimal Hyperplane Becomes

the True One
As discussed, the MEMPM derives the worst-case Bayes optimal hyperplane,
thus it is interesting to dig out on what conditions the worst-case optimal
one changes into the true optimal one.
In the following we demonstrate two propositions: the rst is that when
data are assumed under some distributions, e.g. Gaussian distribution, the
MEMPM leads to the Bayes optimal classier; the second is that when applied
into high-dimensional classication tasks, the MEMPM can be adapted to
converge into the true Bayes optimal classier under the Lyapunov condition.
To introduce the rst proposition, we begin with assuming data distribution as a Gaussian distribution.
Assuming x N (x, x ) and y N (y, y ), Eq.(3.3) becomes:
40
inf
xN (x, x )
P r{wT x b} = P rxN (x, x ) {wT x b}

b wT x
= P r N (0, 1)
wT x w

b wT x
= 1
wT x w

b + wT x
,
=
wT x w
(3.24)
where (z) is the cumulative distribution function for the standard normal
Gaussian distribution dened as:
2
z
1
s
ds.
(z) = P r{N (0, 1) z} =
exp
2
2
Due to the monotonic property of (z), we can further write Eq.(3.24) as:
b + wT x 1 () wT x w .
Constraint Eq.(3.4) can be reformulated to a similar form. The optimization
Eq.(3.2) is thus changed as:
{ + (1 )} ,
b + wT x 1 () wT x w ,

b wT y 1 () wT y w .
max
,,w=0,b
s.t.
(3.25)
(3.26)
The above optimization is nearly the same as Eq.(3.2) subject to the con1
straints
of Eqs.(3.3) and (3.4) except that, () is equal to (), instead
of
1 . Thus, it can be similarly solved based on the Sequential Biased
Minimax Probability Machine method.
On the other hand, the Bayes optimal hyperplane corresponds to the one,
wT z = b, which minimizes the Bayes error:
min
w=0,b
P rxN (x, x ) {wT x b} + (1 )P ryN (y, y ) {wT y b} (3.27)
The above is exactly the upper bound of + (1 ). From Lemma 3.4 we

can know that Eq.(3.26) will eventually become equalities. Traced back to
Eq.(3.24), the equalities imply that and will achieve their upper bounds
respectively. Therefore, with the Gaussian distribution assumption on data,
the MEMPM derives the optimal Bayes hyperplane.
We propose Proposition 3.5 to extend the above analysis to general distribution assumptions.
41
Proposition 3.5. If the distribution of the normalized random variable

wT x wT x
,
wT x w
denoted as N S, is independent of w, as the case in Gaussian distribution,
the similar MEMPM version as in Gaussian distribution assumption will be
easily derived, except that (z) is changed as P r{N S(0, 1) z}. In such
case, minimizing the Bayes error bound will exactly minimize the true Bayes
error.
Before presenting Proposition 3.7, we rst introduce the Central Limit
Theorem under the Lyapunov condition [5].
Theorem 3.6. Let xn be a sequence of independent random variables dened
on the same probability space. Assume that xn has nite expected value n
n

and nite standard deviation n . We dene s2n =
i2 . Assume that the
third central moment
rn3
n

i=1
i=1
E(|xn n | ) is nite for every n, and that

3
lim (rn /sn ) = 0 (This is the Lyapunov condition).

n
xn converges towards a Gaussian distribution.
The sum S n = x1 + ... +
One interesting nding directly elicited from the above Central Limit
Theorem is that, if the component variable xi of a given n-dimensional random variable x satises the Lyapunov condition, the sum of weighted component variables xi , 1 i n, namely, wT x tends to be a Gaussian distribution, as n grows.5 This shows that, under the Lyapunov condition, when
the dimension n grows, the hyperplane derived by MEMPM with Gaussian
assumption tends to be the true Bayes optimal hyperplane. In this case, the
MEMPM using 1 (),
the inverse function of the normal cumulative distribution, instead of /(1 ), will converge to the true Bayes optimal
decision hyperplane in the high-dimensional space. We summarize the analysis into Proposition 3.7.
Proposition 3.7. If the component variable xi of a given n-dimensional random variable x satises the Lyapunov condition, the MEMPM hyperplane derived by using 1 () the inverse function of normal cumulative distribution,
will converge to the true Bayes optimal one.
The underlying justications in the above two propositions root in the
fact that the generalized MPM is exclusively determined by the rst and second moments. These two propositions actually emphasize the dominance of
the rst and second moments in representing data. More specically, Proposition 3.5 hints that the distribution is only decided by up to the second
5
Some techniques such as Independent Component [8] can be applied to decorrelate the dependence among random variables beforehand.
42
moment. The Lyapunov condition in Proposition 3.7 also implies that the
second order moment dominates the third order moment in the long run. It
also deserves attention that with the xed mean and covariance, the distribution of Maximum Entropy Estimation is the Gaussian distribution [14]. This
would once again suggest the usage of 1 () in the high-dimensional space.
3.2.6 Geometrical Interpretation
In this section, we rst provide a parametric solving method for BMPM, then
demonstrate that this parametric method actually enables a nice geometrical
interpretation for both BMPM and MEMPM.
3.2.6.1 A Parametric Method for BMPM
According to the parametric method, the fractional function can be iteratively optimized in two steps [26]:
Step 1. Find w by maximizing f (w) g(w) in the domain A, where R
is the newly introduced parameter.
Step 2. Update by f (w)/g(w).
The iteration of the above two steps will guarantee to converge to the local
maximum which is also the global maximum in our problem. In the following,
we adopt a method to solve the maximization problem in Step 1. Replacing
f (w) and g(w), we expand the optimization problem as:

max 1 (0 ) wT y w wT x w , s.t. wT (x y) = 1. (3.28)

w=0
Maximizing Eq.(3.28) is equivalent to minw (0 ) wT y w + wT x w

under the same constraint. By writing w = w0 + F u, where w0 = (x y)/
x y 22 and F Rn(n1) is an orthogonal matrix whose columns span
the subspace of vectors orthogonal to x y, an equivalent form (a factor 1/2
over each term has been dropped) to remove the constraint can be obtained:

2
(0 )2
1/2
1/2
2
2
min
+ x (w0 + F u)2 + +
y (w0 + F u)2 ,
u,>0,>0
(3.29)
where , R. This optimization form is very similar to the one in Minimax
Probability Machine [15] and can also be solved by using an iterative leastsquares approach.
3.2.6.2 A Geometrical Interpretation for BMPM and MEMPM
The parametric method actually enables a nice geometrical interpretation of
BMPM and MEMPM in a fashion similar to that of MPM in [16]. Similarly,
43
we assume x = y for the meaningful classication and also assume that x

and y are positive denite for the purpose of simplicity.
By using the 2-norm denition of a vector z : z2 = max{uT z : u2
1}, we can express Eq.(3.28) in its dual form:

:= min max
uT x 1/2 w + (0 )v T y 1/2 w + (1 wT (x y))
w=0 u,v
s.t. u2 1, v2 1 .
We change the order of the min and max operators and consider the min:
min {uT x 1/2 w + (0 )v T y 1/2 w + (1 wT (x y))}

,
if x x 1/2 u = y + (0 ) y 1/2 v;
=
, otherwise.
w=0
Thus, the dual problem can be further changed to:

max : u2 1, v2 1, x x 1/2 u = y + (0 ) y 1/2 v. (3.30)
,u,v
By dening := 1/ we rewrite the dual problem as:

min
,u,v
: x x 1/2 u = y + (0 ) y 1/2 v, u2 , v2 . (3.31)
When the optimum is attained, we have

= x 1/2 w 2 + (0 ) y 1/2 w 2 = 1/ .
(3.32)
We consider each side of Eq.(3.31) as an ellipsoid centered at the means

x and y and shaped by the weighted covariance matrices x and (0 ) y
respectively:
Hx () = {x = x + x 1/2 u : u2 },
(3.33)
Hy () = {y = y + (0 ) y
(3.34)
1/2
v : v2 }.
The above optimization involves nding a minimum for which two ellipsoids intersect. For the optimum , these two ellipsoids would be tangent to
each other. We further note that, according to Lemma 3.4, at the optimum,
, which is maximized via a series of the above procedures, would satisfy
1 = x 1/2 w 2 + (0 ) y 1/2 w 2 = = 1/ ,
= 1 .
(3.35)
(3.36)
This means that the ellipsoid for the class y nally changes to the one
centered at y, whose Mahalanobis distance to y is exactly equal to (0 ).
Moreover, the ellipsoid for the class x would be the one centered at x and
44
tangent to the ellipsoid for the class y. In comparison, for MPM, two ellipsoids grow with the same speed (with the same () and ()). On the
other hand, since MEMPM corresponds to solving a sequence of BMPMs,
it similarly leads to a hyperplane tangent to two ellipsoids, which achieves
to minimize the maximum of the worst-case Bayes error. Moreover, it is not
necessarily attained in a balanced way as in MPM, i.e. two ellipsoids do not
necessarily grow with the same speed and hence probably contain the unequal
Mahalanobis distance from their corresponding centers. This is illustrated in
Fig. 3.3.
Fig. 3.3. The Geometrical interpretation of MEMPM and BMPM. Finding

the optimal BMPM hyperplane corresponds to nding the decision plane
(the black dashed line) tangent to an ellipsoid (the inner dashed ellipsoid
on the y side) , which is centered at y, shaped by the covariance y and
whose Mahalanobis distance to y is exactly equal to (0 ) ((0 ) = 1.28
in this example). The worst-case accuracy for x is determined by the
Mahalanobis distance ( = 5.35 in this example), at which an ellipsoid
(centered at x and shaped by x ) is tangent to that (0 ) ellipsoid, i.e. the
outer dahsed ellipsoid on the x side. In comparison, MPM tries to nd out
the minimum equality-constrained , at which two ellipsoids for x and y
intersect (both dotted ellipsoids with = 2.77). For MEMPM, it achieves
a tangent hyperplane in a non-balanced fashion, i.e. two ellipsoids may not
attain the same but are globally optimal in the worst-case setting (see
the solid ellipsoids)
3.3 Robust Version
45
3.3 Robust Version

In the above, the estimates of means and covariance matrices are assumed
reliable. We now consider how the probabilistic framework in Eq.(3.2) changes
against the variation of the means and covariance matrices:
max { + (1 )} ,
,,w=0,b
s.t.
(3.37)
inf
P r{wT x b} , (
x, x ) X ,
(3.38)
inf
P r{wT y b} , (
y, y ) Y ,
(3.39)
x(
x, x )
y(
y , y )
where X and Y are the sets of means and covariance matrices and are the
subsets of RPn+ , where Pn+ is the set of nn symmetric positive semidefinite
matrices.
Motivated by the tractability of the problem and from the statistical view,
a specific setting of X and Y is proposed in [16]. However, they consider the
same variations of the means for two classes, which is easy to handle but less
general. Now, considering the unequal treatment of each class, we propose
the following setting which is in a more general and complete form:

0 ) x1 (
0 ) x2 , x x x0 F x ,
X = (
x, x ) | (
xx
xx

0 ) y 1 (
0 ) y2 , y y y 0 F y ,
yy
yy
Y = (
y , y ) | (
0 , 0x are the nominal means and covariance matrices obtained
where x
through estimating. Parameters x , y , x , and y are positive constants.
The matrix norm is dened as the Frobenius norm: M 2F = Tr(M T M ).
With the assumption that variations of the means for two classes are the
same, the parameters x and y are required equal in [16]. This may enable
the direct usage of the MPM optimization into its robust version. However,
the assumption may not be true in real cases. Moreover, in MEMPM, this
requirement is also not necessary and inappropriate. This will be later demonstrated in the experiment.
By applying the results from [16], we obtain the robust MEMPM as:
max { + (1 )} ,
,,w=0,b

(() + x ) wT ( x0 + x I n )w,
s.t. b + w x

0 (() + y ) wT ( y 0 + y I n )w.
b wT y
T
Analogously, we transform the above optimization problem as:

max
,,w=0
2r ()
+ (1 ) ,
1 + 2r ()
0 ) = 1,
x0 y
s.t. wT (
(3.40)
(3.41)
46

where r () = max
1(()+y )
wT ( y 0 +y I n )w
wT y( x0 +x I n )w

x , 0 , and thus can be
solved by the SBMPM method. The optimal b is therefore calculated by:

0 (( ) + x ) w T ( x0 + x I n )w
b = w T x

0 + (( ) + y ) w T ( y 0 + y In )w .
= w T y
Remarks. Interestingly, if MPM is treated with unequal robust parameters
x and y , it leads to solving an optimization similar to MEMPM, since
() + x will not be equal to () + y . In addition, similar to the robust
MPM, when applied in practice, the specic values of x , y , x and y can
be provided based on the Central Limit Theorem.
3.4 Kernelization
We note that, in the above, the classier derived from MEMPM is given in
a linear conguration. In order to handle nonlinear classication problems,
in this section, we seek to use the kernelization trick [22] to map the ndimensional data points into a high-dimensional feature space Rf , where a
linear classier corresponds to a nonlinear hyperplane in the original space.
Since the optimization of MEMPM corresponds to a sequence of BMPM
optimization problems, this model naturally inherits the kernelization ability of BMPM. We thus in the following mainly address the kernelization of
BMPM.
Ny
x
Assuming training data points are represented by {xi }N
i=1 and {y j }j=1
for the class x and y, respectively, the kernel mapping can be formulated as:
x (x) ((x), (x) ) ,
y (y) ((y), (y) ) ,
where : Rn Rf is a mapping function. The corresponding linear classier in Rf is wT (z) = b, where w, (z) Rf , and b R. Similarly, the
transformed FP optimization in BMPM can be written as:

1 (0 ) wT (y) w

max
, s.t. wT ((x) (y)) = 1. (3.42)
w=0
T
w (x) w
However, to make the kernel work, we need to represent the nal decision
hyperplane and the optimization in a kernel form, K(z 1 , z 2 ) = (z 1 )T (z 2 ),
namely an inner product form of the mapping data points.
3.4 Kernelization
47
3.4.1 Kernelization Theory for BMPM

In the following, we demonstrate that although BMPM possesses a signicantly dierent optimization form from MPM, the kernelization theory proposed in [16] is still viable, provided that suitable estimates for means and
covariance matrices are applied therein.
We rst state a theory similar to Corollary 5 of [16] and prove its validity
in BMPM.
Corollary 3.8. If the estimates of means and covariance matrices are given
in BMPM as:
(x) =
Nx

i (xi ) ,
i=1
(x) = x I n +
(y) =
Ny

j (y j ) ,
j=1
Nx

i ((xi ) (x))((xi ) (x))T ,
i=1
(y) = y I n +
Ny

j ((y j ) (y))((y j ) (y))T ,
j=1
where I n is the identity matrix of dimension n, then the optimal w in problem

Eq.(3.42) lies in the space spanned by the training points.
Proof. Similar to Corollary 5 of [16], we write w = wp + wd , where wp
is the projection of w in the vector space spanned by all the training data
points and wd is the orthogonal component to this span space. It can be
easily veried that Eq.(3.42) changes to maximize the following:

Nx
T
T
T
1(0 ) wT
p
i=1 i ((xi )(x))((xi )(x)) w p +x (w p w p +w d w d )

,

Ny
T
T
T
wT
p
j=1 j ((y j )(y))((y j )(y)) w p +y (w p w p +w d w d )
subject to the constraints of wT

p ((x) (y)) = 1. Since we intend to maximize the fractional form and both the denominator and the numerator are
positive, the denominator needs to be as small as possible and the numerator needs to be as large as possible. This would nally lead to wd = 0. In
other words, the optimal w lies in the vector space spanned by all the training data points. Note that the introduction of x and y actually enables a
direct application of the robust estimates in the kernelization.
According to Corollary 3.8, if appropriate estimates of means and covariance matrices are applied, the optimal w can be written as the linear
combination of training points. In particular, if we obtain the means and
covariance matrices as the plug-in estimates, i.e.
48
(x) =
Nx
1
(xi ) ,
Nx i=1
(y) =
Ny
1
(y j ) ,
Ny j=1
(x) =
Nx
1
((xi ) (x))((xi ) (x))T ,
Nx i=1
(y) =
Ny
1
((y j ) (y))((y j ) (y))T ,
Ny j=1
we can write w as:

w=
Nx

i (xi ) +
i=1
Ny

j (y j ),
(3.43)
j=1
where the coecients i , j R for i = 1, . . . , Nx and j = 1, . . . , Ny .

By simply substituting Eq.(3.43) and four plug-in estimates into Eq.(3.42),
we can obtain the Kernelization Theorem of BMPM.
3.4.2 Notations in Kernelization Theorem of BMPM
Before we present the main kernelization result, we rst introduce the notations. Let {z}N
i=1 denote all N = Nx + Ny data points in the training set
where
z i = xi ,
z i = y iNx ,
i = 1, 2, . . . , Nx ,
i = Nx + 1, Nx + 2, . . . , N.
The element of the Gram matrix K in the position of (i, j) is dened

as K i,j = (z i )T (z j ) for i, j = 1, 2, . . . , N . We further dene Kx and Ky
as the matrices formed by the rst Nx rows and the last Ny rows of K,
respectively, namely,

Kx
.
K :=
Ky
By setting the row average of the K x block and the K x block to zero,
is thus obtained:
the block-row-averaged Gram matrix K

T
x
K x 1Nx k
K
x
K :=
,
=
T
Ky
K y 1N k
y
y
y RNx +Ny are dened as:

x , k
where k
3.4 Kernelization
x ]i :=
[k
49
Ny
Nx

1
y ]i := 1
K(xj , z i ) , [k
K(y j , z i ) .
Nx j=1
Ny j=1
In the above, 1Nx RNx and 1Ny RNy , are dened as:
1i = 1,
1j = 1,
i = 1, 2, . . . , Nx ,
j = 1, 2, . . . , Ny .
Finally, we dene vector formed by the coecients of as:

w = [1 , 2 , . . . , Nx , 1 , 2 , . . . , Ny ]T .
(3.44)
3.4.3 Kernelization Results

Theorem 3.9. [Kernelization Theorem of BMPM] The optimal decision hyperplane of the problem Eq.(3.42) involves solving the Fractional Programming problem:

T
1 (0 ) N1y wT K
y Kyw

,
( ) = max
w=0
1
TK
T
xw
w
K
x
Nx
s.t.
x k
y ) = 1 .
wT (k
The intercept b is calculated as:

b =
wT
kx

1 T T
1 T T
T
( )
w K x K x w = w ky + (0 )
w K K y w ,
Nx
Ny y
where ( ) is obtained when the above equation attains its optimum (w , b ).

For the robust version of BMPM, we can incorporate the variations of the
means and covariances by conducting the following replacements:
1 T T
1 T
w K x K x w wT
K K x + x K)w ,
(
Nx
Nx x
1 T
1 T T
w K y K y w wT
K K y + y K)w ,
(
Ny
Ny y
(0 ) (0 ) + y ,
( ) ( ) + x .
The optimal decision hyperplane can be represented as a linear form in the
kernel space
f (z) =
Nx

i=1
wi K(z, xi ) +
Ny

i=1
wNx +i K(z, y i ) b .
50
3.5 Experiments
In this section, we rst evaluate our model on a synthetic dataset. Then we
compare the performance of MEMPM with that of MPM, on six real-world
benchmark datasets (since MPM is reported comparable to SVM, we do
not perform comparisons with SVM). To demonstrate that BMPM is ideal
for imposing a specied bias in classication, we also implement it on the
Heart-disease dataset. The means and covariance matrices for two classes are
obtained directly from the training datasets by plug-in estimations. The prior
probability is given by the proportion of x data in the training dataset.
3.5.1 Model Illustration on a Synthetic Dataset
To verify that the MEMPM model achieves the minimum Bayes error rate
in the Gaussian distribution, we synthetically generate two classes of twodimensional Gaussian data. As plotted in Fig. 3.4(a), data associated with the
class x are generated with the mean x as [3, 0]T and the covariance matrix x
as [4, 0; 0, 1], while data associated with the class y are generated with the
mean y as [1, 0]T and the covariance matrix y as [1, 0; 0, 5]. The solved
decision hyperplane z1 = 0.333 given by MPM is plotted as the solid line
and the solved decision hyperplane z1 = 0.660 given by MEMPM is plotted
as the dashed line. From the geometrical interpretation, both hyperplanes
should be perpendicular to the z1 axis.
As shown in Fig. 3.4(b), the MEMPM hyperplane exactly represents the
optimal thresholding under the distributions of the rst dimension for two
classes of data, i.e. the intersection point of two density functions. On the
other hand, we nd that the MPM hyperplane exactly corresponds to the
thresholding point with the same error rate for two classes of data, since the
cumulative distribution Px (z1 < 0.333) and Py (z1 > 0.333) are exactly the
same.
3.5.2 Evaluations on Benchmark Datasets
We next evaluate our algorithm on six benchmark datasets. Data for the
Twonorm problem were generated according to [4]. The rest ve datasets
including the Breast, Ionosphere, Pima, Heart-disease, and Vote data were
obtained from UCI machine learning repository [3]. Since handling the missing attribute values is out of the scope of this chapter, we simply remove
instances with missing attribute values in these datasets.
We randomly partition data into 90% training and 10% test sets. The
nal results are averaged over 50 random partitions of data. We compare the
performance of MEMPM and MPM in both the linear setting and Gaussian
kernel setting. The width parameter () for the Gaussian kernel is obtained
3.5 Experiments
51
Fig. 3.4. An evaluation of MEMPM and MPM on a synthetic dataset.

The decision hyperplane derived from MEMPM (the dashed line) exactly
corresponds to the optimal threshholding point, i.e. the intersection point,
while the decision hyperplane given by MPM (the solid line) corresponds
to the point on which two error rates for two classes of data are equal
via cross validations over 50 random partitions of the training set. The experimental results are summarized in Tables 3.1 and 3.2 for the linear kernel
and Guassian kernel respectively.
From the results we can see that our MEMPM demonstrates better performance than MPM in both the linear and Gaussian kernel setting. Moreover,
as observed in these benchmark datasets, the MEMPM hyperplanes are ob-
52

Table 3.1. Lower bound , , and test accuracy compared to MPM in the
linear setting
Performance of MEMPM(%)
Dataset
+ (1 ) Accuracy
Performance of MPM(%)
Accuracy
Twonorm
80.3 0.2 79.9 0.1
80.1 0.1
97.9 0.1 80.1 0.1
97.9 0.1
Breast
77.8 0.8 91.4 0.5
86.7 0.5
96.9 0.3 84.4 0.5
97.0 0.2
Ionosphere
95.9 1.2 36.5 2.6
74.5 0.8
88.5 1.0 63.4 1.1
84.8 0.8
Pima
0.9 0.0 62.9 1.1
41.3 0.8
76.8 0.6 32.0 0.8
76.1 0.6
Heart-disease 43.6 2.5 66.5 1.5
56.3 1.4
84.2 0.7 54.9 1.4
83.2 0.8
82.6 1.3 84.6 0.7
83.9 0.9
94.9 0.4 83.8 0.9
94.8 0.4
Vote
Table 3.2. Lower bound , , and test accuracy compared to MPM in the
Gaussian kernel
Performance of MEMPM(%)
Dataset
+ (1 ) Accuracy
Performance of MPM(%)
Accuracy
Twonorm
91.7 0.2 91.7 0.2
91.7 0.2
97.9 0.1 91.7 0.2
97.9 0.1
Breast
88.4 0.6 90.7 0.4
89.9 0.4
96.9 0.2 89.9 0.4
96.9 0.3
Ionosphere
94.2 0.8 80.9 3.0
89.4 0.8
93.8 0.4 89.0 0.8
92.2 0.4
Pima
2.6 0.1 62.3 1.6
41.4 1.1
77.0 0.7 32.1 1.0
76.2 0.6
Heart-disease 47.1 2.2 66.6 1.4
58.0 1.5
83.9 0.9 57.4 1.6
83.1 1.0
85.1 1.3 84.3 0.7
84.7 0.8
94.7 0.5 84.4 0.8
94.6 0.4
Vote
tained with signicantly unequal and except in the Twonorm set. This
further conrms the validity of our proposition, i.e. the optimal minimax machine is not certain to achieve the same worst-case accuracies for two classes.
For the Twonorm, it is also not an exception. The two classes of data in this
set are generated under the multivariate normal distributions with the same
covariance matrices. In this special case, the intersection point of two density
functions will exactly represent the optimal thresholding point and the one
with the same error rate for each class as well. Another important nding is
that the accuracy bounds, namely + (1 ) in MEMPM and in MPM
are all increased in the Gaussian kernel setting when compared with those
in the linear setting. This shows the advantage of the kernelized probability
machine over the linear probability machine.
In addition, to clearly see the relationship between the bounds and the
test set accuracies (T SA), we plot them in Fig. 3.5. As observed, the test
set accuracies including T SAx (for the class x), T SAy (for the class y), and
the overall accuracies T SA are all greater than their corresponding accuracy
bounds both in MPM and MEMPM. This demonstrates how the accuracy
bound can serve as the performance indicator on future data.
3.5 Experiments
53
Fig. 3.5. Empirical evaluations on bounds and test set accuracies of MEMPM. The
test accuracies including T SAx (for the class x), T SAy (for the class y), and the
overall accuracies T SA are all greater than their corresponding accuracy bounds
both in MPM and MEMPM. This demonstrates how the accuracy bound can serve
as the performance indicator on future data
It is also observed that the overall worst-case accuracies + (1 )

in MEMPM are greater than in MPM both in the linear and Gaussian
settings. This again demonstrates the advantages of MEMPM over MPM.
54
Since the lower bounds keep well with the test accuracies in the above
experimental results, we do not perform the robust version of both models for
the real-world datasets. To see how the robust version works we generate two
classes of Gaussian data. As illustrated in Fig. 3.6, the x data are sampled
(a) Robust MPM and MEMPM with x = y
(b) Robust MPM and MEMPM with x = y

Fig. 3.6. An example in R2 demonstrates the results of robust versions of
MEMPM and MPM. Training points are indicated with black +s for the
class x and magenta s for class y. Test points are represented by blue s
for class x and by green os for the class y. In (a), the robust MEMPM outperforms both MEMPM and the robust MPM. In (b), the robust MEMPM
with x = y outperforms the robust MEMPM with x = y .
from the Gaussian distribution with the mean as [3, 0]T and the covariance
as [1 0; 0 3], while the y data are sampled from another Gaussian distribution
with the mean as [3, 0]T and the covariance as [3 0; 0 1]. We randomly select
3.5 Experiments
55
10 points of each class for training and leave the rest points for test from the
above synthetic dataset. We present the result in the following.
0 and y
0 , covariance matriFirst, we calculate the corresponding means x
0
0
ces x and y and plug them into the linear MPM and the linear MEMPM.
We obtain the MPM decision line (dotted line) with a lower bound (assuming
the Gaussian distribution) being 99.1% and the MEMPM decision line (dashdot line) with a lower bound as 99.7% respectively. However, for the test set
we only obtain the accuracies 93.0% for MPM and 97.0% for MEMPM (see
Fig. 3.6(a)). This obviously violates the lower bound.
Based on our knowledge of the real means and covariance matrices in this
example, we set the parameters as

0 )T x1 (
0 ) = 0.046 ,
x = (
xx
xx

0 )T y1 (
0 ) = 0.496 ,
y = (
yy
yy
x = x x0 F = 1.561 ,
y = y y0 F = 0.972 ,
= max(x , y ) .
We then train the robust linear MPM and the robust linear MEMPM by
these parameters and obtain the robust MPM decision line (dashed line), the
robust MEMPM decision line (solid line), as seen in Fig. 3.6(a). The lower
bounds decrease to 87.3% for MPM and 93.2% for MEMPM respectively,
but the test accuracies increase to 98.0% for MPM and 100.0% for MEMPM.
Obviously, the lower bounds accord with the test accuracies.
Note that in the above, the robust MEMPM also achieves a better performance than the robust MPM. Moreover, x and y are not necessarily
the same. To see the result of MEMPM when x and y are forced to be
the same, we train the robust MEMPM again by setting the parameters as
x = y = as used in MPM. We obtain the corresponding decision line
(dash-dot line) as seen in Fig. 3.6(b). The lower bound decreases to 91.0%
and the test accuracy decreases to 98.0%. The above example indicates how
the robust MEMPM clearly improves over the standard MEMPM when a
bias is incorporated by the inaccurate plug-in estimates and also validates
that x need not be equal to y .
3.5.3 Evaluations of BMPM on Heart-disease Dataset
To demonstrate the advantages of the BMPM model in dealing with biased
classications, we implement BMPM on the Heart-disease dataset, where
dierent treatments for dierent classes are necessary. The x class is associated with data with heart diseases, whereas the y class corresponds to data
without heart diseases. Obviously, a bias should be considered for x, since
misclassication of an x case into the opposite class would delay the therapy
56
and is more risky than the other way round. Similarly, we randomly partition data into 90% training and 10% test sets. Also, the width parameter
() for the Gaussian kernel is obtained via cross validations over 50 random
partitions of the training set. We repeat the above procedures 50 times and
report the average results.
By intentionally varying 0 from 0 to 1, we obtain a series of test accuracies, including the x accuracy T SAx , the y accuracy T SAy for both the
linear and Gaussian kernels. For simplicity, we denote the x accuracy in the
linear setting as T SAx (L), while others are similarly dened.
The results are summarized in Fig. 3.5. Four observations are worth highlighting. First, in both linear and Gaussian kernel settings, the smaller 0 ,
the higher the test accuracy for x. This indicates a bias can be indeed embedded in the classication boundary for the important class x by specifying a
relatively smaller 0 . In comparison, MPM forces an equal treatment on each
class and thus is not suitable for biased classication. Second, the test accuracies for y and x are strictly lower bounded by 0 and . This shows how a bias
can be quantitatively, directly and rigorously imposed towards the important
class x. Note that again, for other weight-adapting-based biased classiers,
the weights themselves lack accurate interpretations and thus cannot rigorously impose a specied bias, i.e. they would try for dierent weights for a
specied bias. Third, when given a prescribed 0 , the test accuracy for x and
its worst-case accuracy in the Gaussian kernel setting are both increased
compared to the corresponding accuracies in the linear setting. Once again,
this demonstrates the power of the kernelization. Fourth, we note that 0
actually contains an upper bound which is around 90% for the linear BMPM
in this dataset. This is reasonable. Observed from Eq.(3.11), the maximum
0 denoted as 0 max is decided by setting = 0, i.e.
(0 max ) = max
w=0
wT
yw
s.t. wT (x y) = 1 .
(3.45)
It is interesting noting that when 0 is set to zero, the test accuracies for
y in the linear and Gaussian settings are both around 50% (see Fig. 3.7(b)).
This seeming irrationality is actually reasonable. We will discuss this in
the next section.
3.6 How Tight Is the Bound?

A natural question for MEMPM is how tight is the worst-case bound. In this
section, we present a theoretical analysis in addressing this problem.
In Marshall and Olkin Theory, if we dene S = {wT y b}, the theorem
is changed to:
sup
y{y, y }
P r{wT y b} =
1
,
1 + d2
with d2 =
inf (y y)T 1
y (y y) .
wT yb
57
Fig. 3.7. Bounds and real accuracies for BMPM in Heart-disease dataset.
With 0 varying from 0 to 1, the real accuracies are lower bounded by the
worst-case accuracies. In addition, (G) is above (L), which shows the
power of the kernelization
Looking into the above equation and Eq.(3.4), for a given hyperplane
{w, b} we can easily obtain:
=
d2
.
1 + d2
(3.46)
Moreover, in [16], a simple closed-form expression for the minimum distance d is derived:
58
d2 =
inf (y y)T y 1 (y y) =
wT yb
max((b wT y), 0)
.
wT y w
(3.47)
It is easy to see that when the decision hyperplane (w, b) passes the center
y, d would be equal to 0 and the worst-case accuracy would be 0 according
to Eq.(3.46).
However, if we consider the Gaussian data (which we assume as y data)
in Fig. 3.8, a vertical line approximating y would achieve about 50% test
accuracy. The large gap between the worst-case accuracy and the real test
accuracy seems strange. In the following, we construct an example of onedimensional data to show the inner rationality of this observation. We attempt to provide the worst-case distribution containing the given mean and
covariance, while a hyperplane passing its mean achieves a real test accuracy
of zero.
Fig. 3.8. Theoretical comparison between the worst-case accuracy and the
real test accuracy for the Gaussian data in Fig. 3.10(a)
Consider the one-dimensional data y consisting of N 1 observations

with
values as m and one single observation with the value as N + m. If we

calculate the mean and the covariance, we obtain:
y =m+ ,
N
N 1 2
.
y =
N
When N goes to innity, the above one-dimensional data have the mean as m
and the covariance as . In this extreme case, a hyperplane passing the mean
will achieve a zero test accuracy which is exactly the worst-case accuracy
59
given the xed mean and covariance as m and respectively. This example
demonstrates the inner rationality of the minimax probability machines.
To further examine the tightness of the worst-case bound in Fig. 3.9(a),
we vary from 0 to 1 and plot against the real test accuracy that a vertical
Fig. 3.9. Three two-dimensional data with the same means and covariances but
with dierent skewness. The worst-case accuracy bound of (a) is tighter than that
of (b) and looser than that of (c)
line classies the y data by using Eq.(3.46). Note that the real accuracy can
be calculated as (z d). This curve is plotted in Fig. 3.10.
Fig. 3.10. Three two-dimensional data with the same means and covariances but with dierent skewness. The worst-case accuracy bound of (a) is
tighter than that of (b) and looser than that of (c)
Observed from Fig. 3.9, the smaller the worst-case accuracy, the looser it
is. On the other hand, if we skew the y data towards the left side, while simul-
60
taneously maintaining the mean and covariance unchanged (see Fig. 3.9(b)),
even a bigger gap will be generated when is small; analogically, if we skew
the data towards the right side (see Fig. 3.9(c)), a tighter accuracy bound will
be expected. This nding would mean that only adopting up to the second
order moments may not achieve a satisfactory bound. In other words, for a
tighter bound, higher order moments such as skewness need to be considered. This problem of estimating a probability bound based on moments is
presented as the (n, k, )-bound problem, which means nding the tightest
bound for n-dimensional variable in the set based on up to the k-th moments. Unfortunately, as proved in [24], it is NP-hard for (n, k, Rn )-bound
problems with k 3. Thus tightening the bound by simply scaling up the
moment order may be intractable in this sense. We may have to exploit other
statistical techniques to achieve this goal. Certainly, this deserves a closer
examination in the future.
3.7 On the Concavity of MEMPM

We address the issue of the concavity on the MEMPM model in this section. We will demonstrate that although MEMPM cannot generally guarantee its concavity, there is strong empirical evidence showing that many realworld problems demonstrate reasonable concavity in MEMPM. Hence, the
MEMPM model can be solved successfully by standard optimization methods, e.g. the linear search method proposed in this chapter.
We rst present a lemma on BMPM.
Lemma 3.10. The optimal solution for BMPM is a strictly and monotonically decreasing function with respect to 0 .
Proof. Let the corresponding optimal worst-case accuracies on x be 1 and
2 respectively, when 0 1 and 0 2 are set as the acceptable accuracy levels
for y in BMPM. We will prove that if 0 1 > 0 2 , then 1 < 2 .
This can be proved by considering the contrary case, i.e. we assume 1
2 . From the problem denition of BMPM, we have:
1 2 = (1 ) (2 )
1 (0 1 ) wT
1 (0 2 ) wT
1 y w1
2 y w2
,(3.48)
T
T
w1 x w1
w2 x w2
where, w1 and w2 are the corresponding optimal solutions which maximize
(1 ) and (2 ) respectively, when 0 1 and 0 2 are specied.
From 0 1 > 0 2 and Eq.(3.48), we have
1 (0 1 ) w1 T y w1
1 (0 2 ) wT
1 y w1
>
(3.49)
wT
w1 T x w1
1 x w1
1 (0 2 ) w2 T y w2
.
(3.50)
w2 T x w2
61
On the other hand, since w2 is the optimal solution of
1 (0 2 ) wT y w
,
max
w=0
wT x w
we have
1 (0 2 ) wT
1 (0 2 ) wT
2 y w2
1 y w1
.
T w
wT
w
w
x 2
x 1
2
1
This is obviously contradictory to Eq.(3.50).

From the sequential solving method of MEMPM, we know that MEMPM
actually corresponds to a one-dimensional line search problem. More specically, it further corresponds to maximizing the sum of two functions, namely,
f1 () + f2 ()6 , where f1 () is determined by the BMPM optimization and
f2 () = . According to Lemma 3.10, f1 () strictly decreases as increases.
Thus it is strictly pseudo-concave. However, generally speaking, the sum of
a pseudo-concave function and a linear function is not necessarily a pseudoconcave function and thus cannot assure that every local optimum is the
global optimum. This can be clearly observed in Fig. 3.10. In this gure, f1
is pseudo-concave in three sub-gures; however, the sum f1 + f2 does not
necessarily lead to a pseudo-concave function.
Nevertheless, there is strong empirical evidence showing that for many
well-behaved real world classication problems, f1 is overall concave, which
results in the concavity of f1 + f2 . This is rst veried by the datasets used
in this chapter. We shift from 0 to the corresponding upper bound and
plot out against in Fig. 3.11. It is clearly observed that in all six datasets
including both kernel and linear cases, the curves of against are overall
concave. This motivates us to look further into the concavity of MEMPM.
As shown in the following, when two classes of data are well-separated, f1
would be concave in the main interest region.
We analyze the concavity of f1 () by imagining that changes from
0 to 1. In this process, the decision hyperplane moves slowly from y to x
according to Eq.(3.46) and Eq.(3.47). At the same time, = f1 () should
decrease accordingly. More precisely, if we denote dx and dy respectively as
the Mahalanobis distances that x and y are from the associated decision
hyperplane with a specied , we can formulate the changing of and as:
k1 (dx )dx ,
+ k2 (dy )dy ,
where k1 (dx ) and k2 (dy ) can be considered as the changing rate of and
when the hyperplane lies dx distance far away from x and dy distance far
6
For simplicity, we assume as 0.5. Since a constant does not inuence the
concavity analysis, the factor of 0.5 is simply dropped.
62
(a) Twonorm
(b) Breast
(c) Ionosphere
(d) Pima
(e) Heart-disease
(f) Vote
Fig. 3.11. The curves of against (f1 ) are all concave-like in the datasets
used in this chapter
63
away from y respectively. We consider the changing of against , namely,

f1
:
f1
=
k1 (dx )dx
.
k2 (dy )dy
If we consider that dx and dy insensitively change against each other or

change with a proportional rate, i.e. dx cdy (c is a positive constant)
as the decision hyperplane moves, the above equation can be further written
as
f1
= c
k1 (dx )
.
k2 (dy )
corresponding 0.25, k2 (dy ) deLemma 3.11. (1) If dy 1/ 3 or the

creases as dy increases. (2) If dx 1/ 3 or the corresponding 0.25,
k1 (dx ) decreases as dx increases.
Proof. Since (1) and (2) are actually very similar statements, we only prove
(1). k2 (d) is actually the rst order derivative of d2 /(1 + d2 ) according to
Eq.(3.46). We consider the rst order derivative of k2 (d) or the second order
2
2
2
2
derivative
of d /(1 + d ). It is easily veried that (d /(1 + d )) 0 when
d 1/ 3. This is also illustrated in Fig. 3.12. According to the denitionof
the second derivative, we immediately obtain the lemma. Note that d 1/ 3
corresponds to 0.25. Thus the condition can be also replaced by 0.25.
In the above procedure, dy , increase and dx , decrease as the hyperplane moves towards x. Therefore, according to Lemma 3.11, k1 (dx ) increases
while k2 (dy ) decreases when , [0.25, 1). This shows that f1
is getting
smaller as the hyperplane moves towards x. In other words, f1
would be
less than 0 and thus is concave when , [0.25, 1). It should be noted
that in many well-separated real world datasets, the optimal and will be
greater than 0.25 with a high possibility, since to achieve good performance,
the worst-case accuracies are naturally required to be greater than a smaller
amount, e.g. 0.25. This is observed in the datasets used in the chapter. All
the datasets except Pima attain their optimums satisfying this constraint.
For Pima, the overall accuracy is relatively lower, which implies that two
classes of data in this dataset appear to largely overlap each other7 .
An illustration can be also seen in Fig. 3.13. We generate two classes of
Gaussian data with x = [0, 0]T , y = [L, 0]T , and x = y = [1, 0; 0, 1].
The prior probability for each data is set as an equal value 0.5. We plot
the curves of f1 () and f1 () + when L is set as dierent values. It is
7
It is observed, even for Pima, the proposed solving algorithm is still successful,
since is approximately linear as shown in Fig. 3.11. Moreover, due to the fact
that the slope of is slightly greater than 1, the nal optimum naturally leads
to achieve its maximum.
64
Fig. 3.12. An illustration of the concavity of MEMPM. Subgure (a)

shows that when two classes of data overlap largely each other, the optimal
solution of MEMPM lies in the small-value range of and which is usually
not concave. (b), (c), and (d) show that when two classes of data are wellseparated, the optimal solutions lie in the region with , [0.25, 1) which
is often concave
observed that when two classes of data largely overlap each other, for example
in Fig. 3.12(a) with L = 1, the optimal solution of MEMPM lies in the
small-value range of and , which is usually not concave. On the other
hand, Fig. 3.12(b), (c), and (d) show that when two classes of data are wellseparated, the optimal solutions lie in the region with , [0.25, 1), which
is often concave.
Note that, in the above, we make an assumption that as the decision hyperplane moves, dx and dy change at an approximately xed proportional
3.8 Limitations and Future Work
65
2
2
Fig. 3.13.
The curve of d /(1 + d ). This function is concave when
d 1/ 3
rate. From the denition of dx and dy , this assumption implies that w,

the direction of the optimal decision hyperplane, is insensitive to . This
assumption does not hold in all cases; however, observed from the geometrical
interpretation of MEMPM, for those data with isotropic or not signicantly
anisotropic x and y , w would be indeed insensitive to .
We summarize the above analysis into the following proposition.
Proposition 3.12. Assuming (1) two classes of data are well-separated and
(2) dx and dy change at an approximately xed proportional rate as the
optimal decision hyperplane (associated with a specied ) moves, the onedimensional line search problem of MEMPM is often concave in the range of
, [0.25, 1) and will often attain its optimum in this range. Therefore the
proposed solving method leads to a satisfactory solution.
Remarks. As demonstrated in the above, although MEMPM is often overall
concave in real world tasks, there exist cases that MEMPM optimization
problem is not concave. This may lead to the case that the solved local
optimum, based on the SBMPM method, is not the global optimum. In these
instances, we may need carefully choose the initial starting point. In addition,
the physical interpretation of as the worst-case accuracy, may make it
relatively easy to choose a suitable initial value. For example, we can set the
initial value by using the information obtained from prior domain knowledge.
3.8 Limitations and Future Work

In this section, we present the limitations and future work.
First, although MEMPM achieves better performance than MPM, its
sequential optimization of Biased Minimax Probability Machine may cost
66
more training time than MPM. In our experiments, MEMPM needs to solve
5 15 BMPM optimizations on the average. Supposing that BMPM is solved
based on Conjugate Gradient Methods (with a worst-case time complexity
in the same order as MPM), MEMPM would be 5 15 times as expensive as
MPM. Although in pattern recognition tasks, especially in o-line classications, eectiveness is often more important than eciency, expensive timecost presents one of the main limitations of the MEMPM model, in particular
for large scale datasets with millions of samples. To solve this problem, one
possible direction is to reduce those redundant points which actually make
less contributions to the classication. In this way, the problem dimension
(in the kernelization) would be greatly decreased and therefore may help in
reducing the computational time required. Another possible direction is to
exploit some techniques to decompose the Gram matrix (as is done in SVM)
and to develop some specialized optimization procedures for MEMPM. Undoubtedly, speeding up the algorithm will be a highly worthy topic in the
future.
Second, as a generalized model, MEMPM actually incorporates some
other variations. For example, when the prior probability () cannot be estimated reliably (e.g. in sparse data), maximizing +, namely the sum of the
accuracies or the dierence between true positive and false positive, would
be considered. This type of approaches is widely used in pattern recognition
eld, e.g. in medical diagnosis [10] and in graph detection, especially line
detection and arc detection, where it is called Vector Recovery Index [9, 17].
Moreover, when there are domain experts at hand, a variation of MEMPM,
namely, the maximization of Cx + Cy may be used, where Cx (Cy ) is the
cost of a misclassication of x (y) obtained from experts. Exploring these
variations in some specic domains is thus a valuable direction in the future
(we actually will discuss these variations as criteria for biased or imbalanced
learning in Chapter 5).
Third, [16] has built up a connection between MPM and SVM from the
perspective of the margin denition, i.e. MPM corresponds to nding the
hyperplane with the maximal margin from the class center. Nevertheless,
some deeper connections need to be investigated, e.g. how is the bound of
MEMPM related to the generation bound of SVM? More recently, [11] and
also the next chapter have disclosed the relationship between them from
either a local or global viewpoint of data. It is particularly useful to look into
these links and explore their further connections in the future.
3.9 Summary
In this chapter, we have proposed a novel global learning model named Minimum Error Minimax Probability Machine. By minimizing the upper bound of
the Bayes error of future data points, our model derives the distribution-free
Bayes optimal hyperplane in the worst-case setting. This thus distinguishes
References
67
itself from the traditional global learning approaches, or more particularly

from traditional Bayes optimal classsifers. More importantly, we have shown
that the worst-case Bayes optimal hyperplane derived by MEMPM becomes
the true Bayes optimal hyperplane, when some conditions are satised, e.g.
when a Gaussian distribution is assumed on data. We have shown that how
to exploit Mercer kernels in this setting to derive a nonlinear classication
boundary. We also have demonstrated that how a robust framework can be
introduced to make solid the foundation of the proposed model. Moreover, we
have demonstrated that this novel model permits an explicit accuracy bound
on future data theoretically and validate this proposition empirically as well.
We have evaluated our algorithms on both synthetic datasets and real-world
benchmark datasets. The performance of MEMPM is demonstrated to outperform MPM, a comparable model with SVM.
References
1. Bazaraa MS (1993) Nonlinear Programming: Theory and Algorithms. New
York, NY: John Wiley & Sons, 2nd edition
2. Bertsekas DP (1999) Nonlinear Programming. Athena Scientic, Belmont,
Massachusetts, 2nd edition
3. Blake CL, Merz CJ(1998) Repository of machine learning databases, University
of California, Irvine, http://www.ics.uci.edu/mlearn/MLRepository.html
4. Breiman L(1997) Arcing Classiers. Technical Report 460, Statistics Department, University of California
5. Chow YS, Teicher H(1997) Probability Theory: Independence, Interchangeability, Martingales. New York, NY: Springer-Verlag, 3rd edition
6. Craven BD (1978) Mathematical Programming and Control Theory. London,
UK: Chapman & Hall
7. Craven BD (1988) Fractional Programming, Sigma Series in Applied Mathematics 4. Berlin: Heldermann Verlag
8. Deco G, Obradovic D (1996) An Information-theoretic Approach to Neural
Computing. Heidelberg; New York: Springer-Verlag
9. Dori D, Liu W (1999) Sparse pixel vectorization: An algorithm and its performance evaluation. IEEE Trans. Pattern Analysis and Machine Intelligence
21:202215
10. Grzymala-Busse JW, Goodwin LK, Zhang X (2003) Increasing sensitivity of
preterm birth by changing rule strengths. Pattern Recognition Letters 24:903
910
11. Huang K, Yang H, King I, Lyu MR (2004) Learning large margin classiers
locally and globally. In The 21st International Conference on Machine Learning
(ICML-2004)
12. Huang K, Yang H, King I, Lyu MR, Chan L (2003) Biased minimax probability
machine for medical diagnosis. In the Eighth International Symposium on
Articial Intelligence and Mathematics
13. Ibaraki T (1981). Solving mathematical programming problems with fractional
objective functions In S. Schaible and W. T. Ziemba., editors, Generalized
68
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
References
Concavity in Optimization and Economics. New York, NY: Academic Press
441472
Keysers D, Och FJ, Ney H(2002) Maximum entropy and Gaussian models
for image object recognition. In Proceedings of the 24th DAGM Symposium,
Lecture Notes in Computer Science. Heidelberg: Springer-Verlag, LNCS 2449:
498506
Lanckriet GRG, Ghaoui LE, Bhattacharyya C, Jordan MI (2001) Minimax
(NIPS)
Lanckriet GRG, Ghaoui LE, Bhattacharyya C, Jordan MI (2002) A robust
3:555582
Liu W, Dori D (1997) A protocol for performance evaluation of line detection
algorithms. Machine Vision and Application 9:240250
Maloof MA, Langley P, Binford TO, Nevatia R, Sage S (2003) Improved rooftop
detection in aerial images with machine learning. Machine Learning 53:157191
Mangasarian Olvi L (1994) Nonlinear Programming. Philadelphia: Society for
Industrial and Applied Mathematics
Marshall AW, Olkin I (1960) Multivariate Chebyshev inequalities. Annals of
Moulin Herv
e (1995) Cooperative Microeconomics: a game-theoretic introduction. Princeton, NJ: Princeton University Press
M
uller KR, Mika S, R
atsch G, Tsuda K, Sch
o1kopf B (2001) An introduction
to Kernel-based Learning Algorithms. IEEE Transactions on Neural Networks
12:181201
Osuna E, Freund R, Girosi F (1997) Support Vector Machines: Training and
Applications. Technical Report AIM-1602, Cambridge, MA: The MIT Press
Popescu I, Bertsimas D (2001) Optimal inequalities in probability theory: A
convex optimization approach. Technical Report TM62, INSEAD
Schaible S (1977) Fractional programming. Zeitschrift f
ur Operational Research, Serie A 27(1):3954
Schaible S (1995) Fractional programming. In R. Horst and P. M. Pardalos,
editors, Handbook of Global Optimization, Nonconvex Optimization and its
Applications. Dordrecht,Boston,London: Kluwer Academic Publishers 495608
Sch
olkopf B, Smola A(2002) Learning with Kernels. Cambridge, MA: The MIT
Press
4
Learning Locally and Globally: Maxi-Min
Margin Machine
The proposed MEMPM model obtains the decision hyperplane by using only
global information, e.g. the mean and covariance matrices. However, although
these moments can be more reliably obtained than estimating the distribution, they may still be inaccurate in many cases, e.g. when the data are very
sparse.
Recently, local learning methods, especially large margin classiers [19]
have attracted much interest in the community of machine learning and pattern recognition. Support Vector Machine (SVM) [25], the most famous one
of them, represents a state-of-the-art classier. The essential point of SVM
is to nd a linear separating hyperplane, which achieves the maximal margin among dierent classes of data. Furthermore, one can extend SVM to
build nonlinear separating decision hyperplanes by exploiting kernelization
techniques.
These methods do not try to summarize any global information beforehand, but to focus on obtaining the decision hyperplane in a local way. For
example, in SVM the decision boundary is exclusively determined by some
critical points which are called support vectors, whereas all other points are
totally irrelevant to this hyperplane. Although this scheme is both theoretically and empirically demonstrated to be powerful, it actually discards the
global information of data.
An illustration example can be seen in Fig. 4.1. In this gure, the classication boundary is intuitively observed to be mainly determined by the
dotted axis, i.e. the long axis of the y data (represented by s) or the short
axis of the x data (represented by s). Moreover, along this axis, the y data
are more possible to scatter than the x data, since y contains a relatively
larger variance in this direction. Noting this global fact, a good decision
hyperplane seems reasonable to lie closer to the x side (see the dash-dot line).
However, SVM ignores this kind of global information, i.e. the statistical
trend of data occurrence: the derived SVM decision hyperplane (the solid
70
4 Learning Locally and Globally: Maxi-Min Margin Machine
line) lies unbiasedly right in the middle of two local points (the support
vectors)1 .
Fig. 4.1. A decision hyperplane with considerations of both local

and global information
Aiming to construct classiers both locally and globally, we propose the

Maxi-Min Margin Machine (M4 ) in this chapter. We will attempt to combine
the local learning into the global information, i.e. the covariance information,
which can represent the data trend. Moreover, as this model also contains
the properties of local learning, it will naturally neutralize the impact when
the global information is inaccurate.
As we show later, one critical contribution of this novel model is that
M4 actually presents a unied model of SVM and another recently-proposed
promising model Minimax Probability Machine (MPM) [11]. Moreover, based
on our proposed local and global view of data, another popular model, Fisher
Discriminant Analysis (FDA) [4] can also be interpreted as its special case.
Another good feature of the M4 model is that it can be cast as a sequential Conic Programming problem [17], or more specically, a sequential
Second Order Cone Programming (SOCP) problem [12, 15, 10], which thus
can be practically solved in polynomial time. In addition, with incorporating
the global information, a reduction method is proposed for decreasing the
computation time of this new model.
The third important feature of our proposed model is that the kernelization methodology is also applicable for this formulation. This thus generalizes
the linear M4 to a more powerful classication approach which can derive
nonlinear decision boundaries.
The rest of this chapter is organized as follows. In the next section, we
introduce the M4 model in detail, including its model denition, the geometri1
This gure has appeared earlier in Chapter 2. However, for the purpose of
self-containing for each chapter, we still present it here.
71
cal interpretation, connections with other models, and the associated solving
methods. In Section 4.2, we derive a generation bound for the M4 model. In
Section 4.3, we develop a reduction method to remove redundant points for
decreasing the computational time. In Section 4.4, we exploit the kernelization trick to extend M4 to nonlinear classication tasks. In Section 4.5, we
evaluate this novel model on both synthetic datasets and real world benchmark datasets. In Section 4.6, we make discussions on the M4 model and also
present future work. Finally, we conclude this chapter in Section 4.7. This
work can be also seen in [5] [7] for a short version.

In the following, we rst, for the purpose of clarity, divide M4 into separable and nonseparable categories, and then introduce the corresponding
hard-margin M4 and soft-margin M4 sequently. In this section, we will also
establish the connections of the M4 model with other large margin classiers including SVM, MPM, FDA and Mininum Error Minimax Probability
Machine (MEMPM) [6].
4.1.1 Separable Case
Assuming the classication samples are separable, we rst introduce the
model denition and the geometrical interpretation. We then transform the
model optimization problem into a sequential SOCP problem and discuss the
detailed optimization method.
4.1.1.1 Problem Denition
Only two-category classication tasks are considered in this chapter. Let a
training dataset contain two classes of samples represented by xi Rn and
y j Rn respectively, where i = 1, 2, . . . , Nx , j = 1, 2, . . . , Ny . The basic task
here can be informally described to nd a suitable hyperplane f (z) = wT z+b
separating two classes of data as robustly as possible (w Rn \{0}, b R,
and wT is the transpose of w). Future data points z for which f (z) 0 are
then classied as the class x; otherwise, they are classied as the class y.
The formulation for M4 can be written as:
max
,w=0,b
(wT xi + b)
, i = 1, 2, . . . , Nx ,
s.t.
wT x w
(wT y j + b)
, j = 1, 2, . . . , Ny ,
wT y w
(4.1)
(4.2)
(4.3)
72
where x and y refer to the covariance matrices of the x and the y data,
respectively.
This model tries to maximize the margin dened as the minimum Mahalanobis distance for all training samples,while simultaneously classifying all
the data correctly. Compared to SVM, M4 incorporates the data information
in a global way; namely, the covariance information of data or the statistical
trend of data occurrence is considered, while SVMs, including l1 -SVM [27]
and l2 -SVM [24] (lp -SVM means the p-norm distance-based SVM) [19],
simply discard this information or consider the same covariance for each
class.
4.1.1.2 Geometrical Interpretation
A geometrical interpretation of M4 can be seen in Fig. 4.2. In this gure, the
Fig. 4.2. A geometric interpretation of M4 . The M4 hyperplane corresponds to the tangent line (the solid line) of two small dashed ellipsoids
centered at the support vectors (the local information) and shaped by the
corresponding covariances (the global information). It is thus more reasonable than SVM (the dotted line)
x data are represented by the inner ellipsoid on the left side with its center
as x0 , while the y data are represented by the inner ellipsoid on the right
side with its center as y 0 . It is observed that these two ellipsoids contain
unequal covariances or risks of data occurrence. However, SVM does not
consider this global information: its decision hyperplane (the dotted line) is
located unbiasedly in the middle of two support vectors (lled points). In
comparison, M4 denes the margin as a Maxi-Min Mahalanobis distance,
which thus constructs a decision plane (the solid line) with considerations
of both the local and global information: the M4 hyperplane corresponds to
73
the tangent line of two dashed ellipsoids centered at the support vectors (the
local information) and shaped by the corresponding covariances (the global
information).
4.1.1.3 Optimization Method
In the following, we propose the optimization method for the M4 model. We
will demonstrate that the above problem can be cast as a sequential Conic
Programming problem, or more specically, a sequential SOCP problem.
Our strategy is based on the Divide and Conquer technique. One may
note that in the optimization problem of M4 , if is xed to a constant n , the
problem is exactly changed to conquer the problem of checking whether
the constraints of Eqs.(4.2) and (4.3) can be satised. Moreover, as will be
demonstrated shortly, this checking procedure can be stated as an SOCP
problem. Thus the problem now becomes that how is set, which we can
use divide to handle: if the constraints are satised, we can increase n
accordingly; otherwise, we decrease n .
We detail this solving technique in the following two steps:
(1) Divide: Set n = (0 + m )/2, where 0 is a feasible , m is an infeasible
, and 0 m .
(2) Conquer: Call the Modied Second Order Cone Programming (MSOCP)
procedure elaborated in the following to check whether n is a feasible .
If yes, set 0 = n ; otherwise, set m = n .
In the above, if a satises the constraints of Eqs.(4.2) and (4.3), we call it
a feasible ; otherwise, we call it an infeasible . These two steps are iterated
until |0 m | is less than a small positive value.
We propose the following Theorem 4.1 showing that the MSOCP procedure, namely, the checking problem with xed to a constant n , is solvable
by casting it as an SOCP problem.
Theorem 4.1. The problem of checking whether there exist a w and a b
satisfying the following two sets of constraints Eqs.(4.4) and (4.5) can be
transformed as an SOCP problem which can be solved in polynomial time,
(wT xi + b) n wT x w, i = 1, . . . , Nx ,
(4.4)

(wT y j + b) n wT y w, j = 1, . . . , Ny .
(4.5)
Proof. Introducing dummy variables , we rewrite the above checking problem as an equivalent optimization problem:
max
w=0,b,
Nx +Ny
{ min
k=1
s.t. (wT xi + b) n
k}
wT x w i ,

(wT y j + b) n wT y w j+Nx ,
74
where i = 1, . . . , Nx and j = 1, . . . , Ny .
By checking whether the minimum k at the optimum point is positive,
we can know whether the constraints of Eqs.(4.2) and (4.3) can be satised.
If we go further, we can introduce another dummy variable and transform
the above problem into an SOCP problem:
max
w=0,b, ,
wT x w i ,

(wT y j + b) n wT y w j+Nx ,
s.t. (wT xi + b) n
k ,
where i = 1, . . . , Nx , j = 1, . . . , Ny , and k = 1, . . . , Nx + Ny . By checking

whether the optimal is greater than 0, we can immediately know whether
there exist a w and a b satisfying the constraints of Eqs.(4.2) and (4.3).
Moreover, the above optimization is easily veried to be the standard SOCP
form, since the optimization function is a linear form and the constraints are
either linear or the typical second order conic constraints.
Remarks. In practice, many SOCP programs, e.g. Sedumi [20], provide
schemes to directly handle the above checking procedure. It thus need not
introduce dummy variables as what we have done in the proof.
We now analyze the time complexity of M4 . As indicated in [12], if the
SOCP is solved based on interior-point methods, it contains a worst-case
complexity of O(n3 ). If we denote the range of feasible s as L = max min
and the required precision as , then the number of iterations for M4 is
log(L/) in the worst case. Adding the cost of forming the system matrix
(constraint matrix) which is O(N n3 ) (N represents the number of training
points), the total complexity would be O(log(L/)n3 +N n3 ) O(N n3 ) which
is relatively large but can still be solved in polynomial time2 .
4.1.2 Connections with Other Models
In this section, we establish connections between M4 and other models. We
show that SVM and MPM are actually special cases of our model. Moreover,
FDA can be interpreted and extended according to our local and global views
of data.
4.1.2.1 Connection with Minimax Probability Machine
If one expands the constraints of Eq.(4.2) and adds all of them together, one
can immediately obtain the following equation:
2
Note that the system matrix needs to be formed only once.
wT
Nx

75
xi + Nx b Nx wT x w wT x + b wT x w , (4.6)
i=1
where x denotes the mean of the x training data.

Similarly, from Eq.(4.3) one can obtain:
(wT
Ny

y j + Ny b) Ny wT y w
j=1
(wT y + b)
wT y w ,
(4.7)
where y denotes the mean of the y training data.

Adding Eqs.(4.6) and (4.7), one can obtain:
max
,w=0
s.t.
wT (x y) ( wT x w + wT y w) .
(4.8)
The above optimization is exactly the MPM optimization [11]. Note, however, that the above procedure cannot be reversed. This means that MPM is
a special case of M4 .
Remarks. In MPM, since the decision is completely determined by the global
information, namely, the mean and covariance matrices [11]3 , to assure an accurate performance the estimates of mean and covariance matrices need to
be reliable. However, it cannot always be the case in real world tasks. On
the other hand, M4 seems to solve this problem in a natural way, because
the impact caused by inaccurately estimated mean and covariance matrices
can be neutralized by utilizing the local information, namely by satisfying
those constraints of Eqs.(4.2) and (4.3) for each local data point. This is also
demonstrated in the later experiment.
4.1.2.2 Connection with Support Vector Machine
If one assumes x = y = , the optimization of M4 can be changed as:
max
,w=0,b
s.t. (wT xi + b) wT w ,
(wT y j + b) wT w ,
where i = 1, . . . , Nx and j = 1, . . . , Ny .
Observing that the magnitude of w will not inuence
the optimization,
without loss of generality, one can further assume wT w = 1. Therefore

the optimization can be changed as:
3
This can be directly observed from Eq.(4.8).
76
min
w=0,b
wT w,
s.t. (wT xi + b) 1 ,
(wT y j + b) 1 ,
(4.9)
(4.10)
(4.11)
where i = 1, . . . , Nx and j = 1, . . . , Ny .
A special case of the above with = I is precisely the optimization of
SVM, where I is the identity matrix.
Remarks. In the above, two assumptions are implicitly made by SVM: One
is the assumption on data orientation or data shape, i.e. x = y = ,
and the other is the assumption on data scattering magnitude or data
compactness, i.e. = I. However, these two assumptions are inappropriate.
We demonstrate this in Figs. 4.3 and 4.4. We assume the orientation and
the magnitude of each ellipsoid represent the data shape and compactness,
respectively, in these gures.
Fig. 4.3. An illustration on that SVM omits the data compactness

information
Fig. 4.3 plots two types of data with the same data orientations but dierent data scattering magnitudes. It is obvious that by ignoring data scattering
SVM is improper to locate itself unbiasedly in the middle of the support vectors (lled points), since x is more possible to scatter on the horizontal axis.
Instead, M4 is more reasonable (see the solid line in this gure). Furthermore,
Fig. 4.4 plots the case with the same data scattering magnitudes but dierent
data orientations. Similarly, SVM does not capture the orientation information. In comparison, M4 grasps this information and demonstrates a more
77
Fig. 4.4. An illustration on that SVM discards the data orientation information
suitable decision plane: M4 represents the tangent line between two small
dashed ellipsoids centered at the support vectors (lled points). Note that
SVM and M4 do not need to achieve the same support vectors. In Fig. 4.4,
M4 contains the above two lled points as support vectors, whereas SVM has
all the three lled points as support vectors.
4.1.2.3 Link with Fisher Discriminant Analysis
FDA, an important and popular method, is used widely in constructing decision hyperplanes and reducing the feature dimensionality. In the following
discussion, we mainly consider its application as a classier. FDA involves
solving the following optimization problem:
|wT (x y)|
max
.
w=0
wT x w + wT y w
Similar to MPM, FDA also focuses on using the global information rather
than considering data both locally and globally. We now show that FDA can
be modied to consider data both locally and globally.
If one changes the denominators in Eqs.(4.2) and (4.3) as wT x w + wT y w ,

the optimization can be changed as:
78
max ,
(4.12)
,w=0,b
(wT xi + b)
s.t.
,
wT x w + wT y w
(4.13)
(wT y j + b)
,
wT x w + wT y w
(4.14)
where i = 1, . . . , Nx and j = 1, . . . , Ny . The above optimization is actually

a generalized case of FDA, which considers data locally and globally. This is
veried as follows.
If one performs the procedure similar to that of Section 4.1.2.1, the above
optimization problem is easily veried to be the following optimization:
max
,w=0,b
(4.15)

s.t. wT (x y) wT x w + wT y w .
One can change Eq.(4.15) as:
|wT (xy)|
,
wT x w+wT y w
which is exactly the
optimization of the FDA (w (x y) is implicitly implied as a positive value

from Eqs.(4.13) and (4.14)).
Remarks. The extended FDA optimization actually focuses on considering
the data orientation, while omitting the data scattering magnitude information. Using the analysis similar to that of Section 4.1.2.2, we can know that
the extended FDA lacks the consideration on the data scattering magnitude.
Its decision hyperplane in the example of Fig. 4.3 coincides with that of
SVM. With respect to the data orientation, it actually uses the average of
covariances for two types of data. As illustrated in Fig. 4.5, the extended
FDA corresponds to the line lying exactly in the middle of the long axes of
the x and y data. This shows that the extended FDA considers the data
orientation partially yet incompletely.
T
4.1.3 Nonseparable Case

In this section, we modify the M4 model to handle the nonseparable case.
We need to introduce slack variables in this case. The optimization of M4 is
changed as:
Nx +Ny

max
C
k ,
(4.16)
,w=0,b,
k=1
(4.17)
s.t. (wT xi + b) wT x w i ,

T
(w y j + b) wT y w j+Nx ,
(4.18)
k 0 ,
79
Fig. 4.5. An illustration on that FDA partly yet incompletely

considers the data orientation
where i = 1, . . . , Nx , j = 1, . . . , Ny , and k = 1, . . . , Nx +Ny . C is the positive

penalty parameter and k is the slack variable which can be considered as
the extent how the training point z k disobeys the margin (z k = xk when
Nx +Ny
1 k Nx ; z k = y kNy when Nx + 1 k Nx + Ny ). Thus k=1
k
can be conceptually regarded as the training error or the empirical error.
In other words, the above optimization achieves maximizing the minimum
margin while minimizing the total training error.
4.1.3.1 Solving Method
As clearly observed, when is xed, the optimization is equivalent to miniNx +Ny
mizing k=1
k under the same constraints. This is once again an SOCP
problem and thus can be solved in polynomial time. We can then update
according to some rules and repeat the whole process until an optimal is
found. This is once again the so-called line search problem. We still adopt
Quadratic Interpolation method to solve this problem, which converges superlinearly to the global optimum if suitable starting points are assigned [1].
Since we have introduced this linear search method in Chapter 3, we simply
omit it here.
In summary, we iterate the following two steps to solve the modied optimization.
Step 1. Generate a new n from three previous 1 , 2 , 3 by using the
Quadratic Interpolation method.
Step 2. Fix = n , perform the optimization based on SOCP algorithms.
Update 1 , 2 , 3 .
80
4.1.4 Further Connection with Minimum Error Minimax

Probability Machine
In this section, we show how the M4 can be connected with Minimum Error Minimax Probability Machine [6], which is a worst-case Bayes optimal
classier and a superset of MPM as well.
If one looks into carefully the optimization
of nonseparable M4 , a more
precise
form is the one replacing k with k wT x w in Eq.(4.17) and
k wT y w in Eq.(4.18). However, this optimization may prove to be a

dicult problem. Nevertheless, we can start from this precise form and derive the connection of M4 with MEMPM.
We reformulate the optimization of Eqs.(4.17) and (4.18) as their precise
forms as follows:
Nx +Ny

max
C
k ,
(4.19)
,w=0,b,
k=1
w xi + b
i ,
s.t.
wT x w
wT y j + b
j+Nx ,
wT y w
k 0 ,
T
(4.20)
(4.21)
(4.22)
where i = 1, . . . , Nx , j = 1, . . . , Ny , and k = 1, . . . , Nx + Ny .
Maximizing Eq.(4.20) contains a similar meaning as minimizing
Nx
+Ny
B
k + 1/2 (B is a positive parameter) in a sense that they both
k=1
attempt to maximize the margin and minimize the error rate. If we conNx
+Ny
sider
k as the residue and regard 1/2 as the regularization term, the
k=1
optimization can be cast into the framework of solving ill-posed problems. 4

According to [24, 26], the above optimization pointed as the Tikhonovs
Variation Method [22] is equivalent to the optimization below refereed to
Ivannovs Quasi-Solution Method [8],in the sense that if one of the methods
for a given value of the parameter (say C) produces a solution {w, b}, then
the other method can derive the same solution by adapting its corresponding
parameter (say A).
4
A trick can be made by assuming 1/2 as a new variable and thus the condition
that the regularization is convex can be satised.
81
Nx +Ny
min
,w=0,b,
k ,
(4.23)
k=1
wT xi + b
i ,
s.t.
wT x w
wT y j + b
j+Nx ,
wT y w
A , k 0 ,
(4.24)
(4.25)
(4.26)
where A is a positive constant parameter.

Now if we expand Eq.(4.24) for each i and add them all together, we can
obtain:
Nx

wT x + b
Nx
Nx
i .
wT x w
i=1
(4.27)
This equation can easily be changed as:

Nx

wT x + b
.
i Nx Nx
wT x w
i=1
(4.28)
Similarly, if we expand Eq.(4.25) for each j and add them all together, we
obtain:
Ny

wT y + b
j+Nx Ny + Ny
.
wT y w
j=1
(4.29)
By adding Eq.(4.28) and Eq.(4.29), we obtain:

N

wT x + b
wT y + b
Ny
k N Nx
.
wT y w
wT x w
k=1
To achieve minimum training error, namely, min,w=0,b,
(4.30)
Nx
+Ny
k=1
k , we
may consider to minimize its lower bound as specied by the right hand side
of Eq.(4.30). Hence in this case should attain its lower bound A, while the
second part should be as large as possible, i.e.

wT y + b
wT x + b
max
(1 )
,
(4.31)
w=0,b
wT y w
wT x w
where is dened as Nx /N and thus 1 denotes Ny /N . If one further
transforms the above to:
82
max
w=0,b
{t + (1 )s},
wT x + b
t,
s.t.
wT x w
wT y + b
s,
wT y w
(4.32)
(4.33)
(4.34)
one can see that the above optimizes a very similar form as the MEMPM
model except that Eq.(4.33) changes to [6]
min {
w=0,b
t2
s2
+
(1
)
}.
1 + t2
1 + s2
In MEMPM, t2 s2 /(1 + t2 )(1 + s2 ) (denoted as ()) represents the worstcase accuracy for the classication of future x (y) data. Thus MEMPM maximizes the weighted accuracy on the future data. In M4 , s and t represent the
corresponding margin which is dened as the distance from the hyperplane
to the class center. Therefore, it represents the weighted maximum margin
machine in this sense. Moreover, since the function of g(u) = u2 /(1 + u2 )
increases monotonically with u, maximizing the above formulae contains a
physical meaning similar to the optimization of MEMPM in some sense.
Remarks. Implicit constraints are contained for the optimization of the
above derived special case of M4 . Empirically, Eq.(4.27) cannot achieve the
equality in the normal case, since Eqs.(4.24) and (4.25) can only achieve
equalities for support vectors. Moreover, the slack variables are usually far
smaller than . This implies we can consider
wT x + b
> = A.
wT x w
Analogously, for y, a similar statement can be obtained. The presence of
these two constraints is essential, since with the constraints the parameter
is involved in the optimization. Moreover, these two constraints also prevent
the circumstance that the decision hyperplane is extremely far away from one
class center, while being very close to the other class center.
4.2 Bound on the Error Rate

In this section, we provide theoretical results on the bound of the error rate
of M4 . We rst borrow the leave-one-out theorem from [13] and [25].
Lemma 4.2. The leave-one-out estimator is almost unbiased.
We then present the generation bound of M4 as the following theorem:
4.2 Bound on the Error Rate
83
Theorem 4.3. If (1) the training set containing N samples is separated by

the decision hyperplane derived by M4 and (2) the mean and covariance matrices are reliably estimated, then the expectation of the probability of the test
error is bounded by the expectation of the minimum of two values: the ratio
m/N and
1
1
+ (1 )
,
1 + d2x
1 + d2y
where m is the number of support vectors, dx and dy are the corresponding Mahalanobis distances from the class centers x and y to the decision
hyperplane, and is prior probability of the x data. Namely,

1
m
1
E[Perror ] E min
.
(4.35)
+
(1
)
,
N 1 + d2x
1 + d2y
Proof. According to Lemma 4.2, to prove E[Perror ] E[ m
N ], we only need
to show that the number of errors by the leave-one-out method does not
exceed the number of support vectors. Actually, this is the case. If we leave a
non-support vector out and then we perform training on the remaining data,
the decision hyperplane will not change, since the decision hyperplane is just
decided by support vectors and the covariance matrices (statistically, one
point will not inuence the covariance of data). Therefore, this non-support
vector will be recognized correctly. Thus the leave-one-out method classies
correctly all the samples that are not support vectors, i.e. the number of the
leave-one-out errors does not exceed

the number of the support vectors.
1
1
We next prove E[Perror ] E min m
. Accor,
+
(1
)
N
1+d2x
1+d2y
ding to [11, 6, 14], if the means and covariances are reliably estimated,
d2x /(1 + dx 2 ) and d2y /(1 + dy 2 ) represent the worst-case rates in recognizing
correctly the x data and y data respectively. Therefore,
1
1
+ (1 )
1 + d2x
1 + d2y
represents the expected maximum error rate, i.e.

m
1
1
E[Perror ] E min
.
+
(1
)
,
N 1 + d2x
1 + d2y
Remarks. Note that the above two items actually represent two meanings
of the M4 model, i.e. minimizing the leave-one-out error presents the contribution by considering the local information from data; on the other hand,
the second item describes the eect by considering the global information
from data. Moreover, if we further examine the second item, dx (dy ) is actually determined by two parts: the Mahalanobis distance from the support
vectors to the corresponding class center x (y) and the margin . This can
be observed in Fig. 4.2. Intuitively, the larger the margin is, the larger dx
and dy are, which leads to a smaller expected test error in the future. This
motivates the margin maximization in the large margin machines.
84
4.3 Reduction
The variables in previous sections are [w, b, 1 , . . . , Nx , . . . , Nx +Ny ], whose
dimension is n + 1 + Nx + Ny . The number of the second order conic constraints is easily veried to be Nx + Ny . This size of the generated constraint
matrix will be a big number and may thus encounter problems in solving
large scale classication tasks. Therefore, we should reduce both the number
of constraints and the number of variables.
Since this problem is caused by the number of the data points, we consider removing some redundant points to reduce both the space and time
complexity. The reduction rule is introduced as follows.
Reduction Rule: Set a threshold [0, 1). In each class, calculate the
Manhalanobis distance di of each point to its corresponding class center. if
d2i /(1 + d2i ) denoted as i is greater than , namely, i , keep this point;
otherwise, remove this point.
The intuition under this rule is that, in general the more discriminant
information the point contains, the further it is from its center (unless it is a
noise point). The inner justication under this rule is from [11]: d2 /(1 + d2 ) is
the worst-case classication accuracy for future data, where d is the minimax
Manhalanobis distance from the class center to the decision hyperplane. Thus
removing those points with small s, namely, d2i /(1 + d2i ) will not aect
the worst-case classication accuracy and will not greatly reduce the overall
performance.
Nevertheless, to cancel the negative impact caused by removing those
points, we add the following global constraint:

wT (x y) ( wT x w + wT y w) .
(4.36)
Integrating the above, we formulate the modied model as follows:

rx +ry

C
max
k + (Nx + Ny rx ry )m
,w=0,b,
k=1
s.t. (w xi + b) ( wT x w) i , i = 1, . . . , rx ,

(wT y j + b) ( wT y w) j+rx , j = 1, . . . , ry ,

wT (x y) ( wT x w + wT y w) m ,
T
m 0,
k 0, k = 1, . . . , rx + ry ,
where, m is the slack variable for the global constraint Eq.(4.36), k are
modied slack variables for the remaining data points, rx is the number of
the remaining points for x, and ry is the number of the remaining points
for y.
4.4 Kernelization
85
Remarks. An interesting observation from the above is that when we set the
reduction threshold to a larger value, or simply to the maximum value 1, the
M4 optimization degrades to the standard MPM optimization. This would
imply that the above modied M4 model contains a worst-case performance
of MPM, if the incorporated local information is useful.
4.4 Kernelization
One may note that in the above, the classier derived from M4 is provided in
a linear conguration. In order to handle nonlinear classication problems,
in this section, we seek to use the kernelization trick [18] to map the ndimensional data points into a high-dimensional feature space Rf , where a
linear classier corresponds to a nonlinear hyperplane in the original space.
The kernel mapping can be formulated as: xi (xi ), y j (y j ),
where i = 1, . . . , Nx , j = 1, . . . , Ny , and : Rn Rf is a mapping function.
The corresponding linear classier in Rf is T (z) = b, where , (z) Rf ,
and b R.
The optimization of M4 in the feature space can be written as:
max
, =0,b
( T (xi ) + b)
s.t.
,
T (x)
( T (y j ) + b)

,
T (y)
(4.37)
i = 1, 2, . . . , Nx ,
j = 1, 2, . . . , Ny .
(4.38)
(4.39)
However, to make the kernel work we need to represent the optimization and
the nal decision hyperplane in a kernel form, K(z 1 , z 2 ) = (z 1 )T (z 2 ),
namely, an inner product form of the mapping data points.
4.4.1 Foundation of Kernelization for M4
In the following, we demonstrate that the kernelization trick indeed works in
M4 , provided suitable estimates of means and covariance matrices are applied
therein.
Corollary 4.4. If the estimates of means and covariance matrices are given
in M4 as the following estimates:
86
(x) =
Nx

i (xi ),
(y) =
i=1
Ny

j (y j ) ,
j=1
(x) = x I n +
Nx

"!
"T
!
i (xi ) (x) (xi ) (x)
,
i=1
(y) = y I n +
Ny

!
"!
"T
j (y j ) (y) (y j ) (y)
,
j=1
where I n is the identity matrix of dimension n, then the optimal in

Eqs.(4.37)(4.39) lies in the space spanned by the training points.
Proof. We write = p + d , where p is the projection of in the vector
space spanned by all the training data points and d is the orthogonal comT
ponent to this span space. By using T
d (xi ) = 0 and d (y j ) = 0, one can
easily verify that the optimization Eqs.(4.37)-(4.39) change to:
max
,{ p , d }=0,b
s.t.
T
p

T
p
,
( T
p (xi ) + b)
N
x
i=1
i ((xj ) (x))((xi )
(x))T
,
p
x ( T
p p
T
d d)
( T
p (y j ) + b)
N
y
j=1
T
j ((y j ) (y))((y j ) (y))T p + y ( T
p p + d d)
where i = 1, . . . , Nx , j = 1, . . . , Ny . Since we intend to maximize the margin

, the denominators in the above two constraints need to be as small as
possible. This would lead to d = 0. In other words, the optimal lies in
the vector space spanned by all the training data points. Note that the above
discussion is assumed in the feature space.
According to Corollary 4.4, if we use the plug-in estimates to approximate
the means and covariance matrices, we can write as a linear combination
form of training data points:
=
Nx

i (xi ) +
i=1
Ny

j (y j ) ,
j=1
where the coecients i , j R, i = 1, . . . , Nx , j = 1, . . . , Ny .

4.4.2 Kernelization Result
We present the kernelization result as the following theorem.
(4.40)
4.4 Kernelization
87
Theorem 4.5. [Kernelization Theorem of M4 ] The optimal decision hyperplane for M4 involves solving the following optimization problem:
max
, =0,b
s.t.
( T K i + b)
T
1
T
Nx K x K x
( T K j+Nx + b)

T
1
T
Ny K y K y
i = 1, 2, . . . , Nx ,
j = 1, 2, . . . , Ny .
Proof. The theorem can easily be proved by simply substituting the plug-in
estimations of means and covariances matrices and Eq.(4.40) into Eqs.(4.38)
(4.39).
The optimal decision hyperplane can be represented as a linear form in
the kernel space:
f (z) =
Nx

i K(z, xi ) +
i=1
Ny

Nx +i K(z, y i ) + b ,
i=1
where and b are the optimal parameters obtained by the optimization

procedure. The notations in the above are dened similar to Chapter 3. However, for an easy reference, we also summarize them in Table 4.1.
Table 4.1. Notations used in Kernelization
Notation
z RNx +Ny
RNx +Ny
K is Gram matrix
x , k
y RNx +Ny
k
1Nx RNx
1Ny RNy
:=
K
z i := xi i = 1, 2, . . . , Nx
z i := y iNx i = Nx + 1, Nx + 2, . . . , Nx + Ny
:= [1 , . . . , Nx , 1 , . . . , Ny ]T
T
K i,j := (z
i ) (z j )
K 1,2
...
K 1,Nx +Ny
K 1,1
K 2,1
K
.
.
.
K
2,2
2,N
+N
x
y
Kx :=
.
.
.
.
.
.
.
.
.
.
.
.
K Nx ,1
K Nx ,2
...
K Nx ,Nx +Ny
K Nx +1,2
. . . K Nx +1,Nx +Ny
K Nx +1,1
K Nx +2,1
K
.
.
.
K
N
+2,2
N
+2,N
+N
x
x
x
y
.
Ky :=
.
.
.
.
.
.
.
.
.
.
.
.
. . . K Nx +Ny ,Nx +Ny
K Nx +Ny ,1 K Nx +Ny ,2

Nx
x ]i := 1
[k
j=1 K(xj , z i ) .
Nx
Ny
1
[ky ]i :=
K(y , z i )
Ny
j=1
1i := 1, i = 1, 2, . . . Nx
1i := 1,
i = 1, 2, . .T. N
y

x
K x 1Nx k
K
x
:=
y
T
K
K y 1Ny k
y
88
4.5 Experiments
In this section, we present the evaluation results of M4 in comparison with
SVM and MPM on both synthetic toy datasets and real world benchmark
datasets. SOCP problems are solved based on the general software named
Sedumi [20, 21]. The covariance matrices are given by the plug-in estimates.
4.5.1 Evaluations on Three Synthetic Toy Datasets
We demonstrate the advantages of our approach in comparison with SVM
and MPM in the following synthetic toy datasets rst.
As illustrated in Fig. 4.6, we generate two types of data with the same
data orientations but dierent data magnitudes in Fig. 4.6 (a), while we generate two types of data with the same data magnitudes but dierent data
orientations in Fig. 4.6 (b). In (a), the x data are randomly sampled from
the Gaussian distribution with the mean as [3.5, 0]T and the covariance as
[3, 0; 0, 4.5], while the y data are randomly sampled from another Gaussian
distribution with the mean and the covariance as [3.5, 0]T and [1, 0; 0, 1.5]
respectively. In (b), the x data are randomly sampled from the Gaussian distribution with the mean as [4, 0]T and the covariance as [1, 0; 0, 5], while
the y data are randomly sampled from another distribution with the mean
and the covariance as [4, 0]T and [1, 0; 0, 5] respectively. Moreover, to generate dierent data orientation, in Fig. 4.6 the y data are rotated anti-clockwise
at the angle of 78 . In both (a) and (b), training (test) data consisting of 120
(250) data points for each class are presented as os (+s) and s (s) for x
and y respectively. Observed from Fig. 4.6, M4 demonstrates its advantages
over SVM. More specically, in Fig. 4.6 (a), SVM discards the information of
the data magnitudes, whose decision hyperplane lies basically in the middle
of boundary points of two types of data, while M4 successfully utilizes this
information, i.e. its decision hyperplane lies closer to the compact class (y
data), which is more reasonable. Similarly, in Fig. 4.6 (b), M4 takes advantage of the information of the data orientation, while SVM simply overlooks
this information, which results in a lot of points incorrectly classied.
In comparison of MPM with M4 , since in the above two datasets the global
information, i.e. the mean and the covariance can be reliably estimated from
data, they achieve similar performance. To see the dierence between M4 and
MPM, we generate another dataset as illustrated in Fig. 4.7, where we intentionally generate a very small number of training data, i.e. only 20 training
points. Similarly, the data are generated under two Gaussian distributions:
the x data are randomly sampled from the Gaussian distribution with the
mean as [3, 0]T and the covariance as [0.5, 0; 0, 8], while the y data are
randomly sampled from another distribution with the mean and the covariance as [4, 0]T and [6, 0; 0, 1] respectively. Training data and test data
are represented using similar symbols to Fig. 4.6. From Fig. 4.7, once again
M4 achieves ideal decision boundary which considers data both locally and
4.5 Experiments
89
(a)
(b)
Fig. 4.6. The rst two synthetic toy examples to illustrate M4 . Training
(test) data consisting of 120 (250) data points for each class are presented as
os (+s) and s (s) for x and y respectively. Subgure (a) demonstrates
that SVM omits the data compactness information and (b) demonstrates
that SVM discards the data orientation information, while M4 achieves
ideal decision boundary which considers data both locally and globally
globally; whereas SVM obtains local boundary just in the middle of the support vectors, which discards the global information, namely the statistical
trend of data occurrence. For MPM, its decision hyperplane is exclusively
dependent on the mean and covariance matrices. Thus we can see that this
hyperplane coincides with the data shape, i.e. the long axis of training data of
x is nearly in the same direction as the MPM decision hyperplane. However,
the estimated mean and covariance are inaccurate due to the small number
of data points. This results in a relatively lower test accuracy as illustrated
in Fig. 4.7(b). In comparison, M4 incorporates the information of the local
points to neutralize the eect caused by inaccurate estimations. The test ac-
90
(a)
(b)
Fig. 4.7. The third synthetic toy example to illustrate M4 . Training (test)
data, consisting of 20 (60) data points for each class are presented as os
(+s) and s (s) for x and y respectively. Subgure (a) demonstrates
the decision boundaries derived from training data, while (b) illustrates
the performance of these hyerplanes on the test set. The M4 achieves ideal
decision boundary which considers data both locally and globally
curacies for the above three toy datasets listed in Table 4.2 also demonstrate
the advantages of M4 .
4.5.2 Evaluations on Benchmark Datasets
We perform evaluations on seven standard datasets. Data for Twonorm problem are synthetically generated according to [3]. The remaining six datasets
are real world data obtained from the UCI machine learning repository [2].
We compared M4 with SVM and MPM engaging with both the linear and
Gaussian kernels. The parameter C for both M4 and SVM was tuned via
4.5 Experiments
91
Table 4.2. Comparisons of classication accuracies between M4 , SVM,

and MPM on the toy datasets
Dataset
Classication accuracy (%)

M4
SVM
MPM
I(%)
98.8
96.8
98.8
II(%)
98.8
97.2
98.8
III(%)
98.3
97.5
95.8
cross validations [9], so was the width parameter in the Gaussian kernel for
all three models. The nal performance results were obtained via the 10-fold
cross validation. Table 4.3 summarizes the evaluation results.
Table 4.3. Comparisons of classication accuracies among M4 , SVM, and MPM
Dataset
Twonorm
Classication accuracy of linear kernel(%) Classication accuracy of Gaussian kernel(%)

M4
SVM
MPM
M4
SVM
MPM
96.5 0.6
95.1 0.7
97.6 0.5
96.5 0.7
96.1 0.4
97.6 0.5
Breast
97.5 0.7 96.6 0.5
96.9 0.8
97.5 0.6 96.7 0.4
96.9 0.8
Ionosphere
87.7 0.8 86.9 0.6
84.8 0.8
94.5 0.4 94.2 0.3
92.3 0.6
Pima
77.7 0.9 77.9 0.7
76.1 1.2
77.6 0.8 78.0 0.5

84.9 1.2
86.5 1.1
76.2 1.2
Sonar
77.6 1.2 76.2 1.1
75.5 1.1
87.3 0.8
Vote
96.1 0.5 95.1 0.4
94.8 0.4
96.2 0.5 95.9 0.6
94.6 0.4
Heart-disease 86.6 0.8 84.1 0.7
83.2 0.8
86.2 0.8 83.8 0.5
83.1 1.0
From the results we observe that M4 achieves the best overall performance. In comparison with SVM and MPM, M4 wins ve cases in the linear
kernel and four in the Gaussian kernel. The evaluations on these standard
bench-mark datasets demonstrate that it is worth considering data both locally and globally, which is emphasized in M4 . Inspecting the dierences
between M4 and SVM, the kernelized M4 appears marginally better than
the kernelized SVM, while the linear M4 demonstrates a distinctive advantage over the linear SVM. This phenomenon may be explained on two hands.
On one hand, this can be explained from the fact that the data points are
very sparse in the kernelized space or feature space (compared with the huge
dimensionality in the Gaussian kernel). Thus the plug-in estimates of the
covariance matrices may not accurately represent the data information in
this case. On the other hand, it is well-known that the kernelization will not
keep the structure information in the feature space. One direct consequence
is that maximizing the margin in the feature space does not necessarily max-
92
imize the margin in the original space [23]. Therefore, without building some
connections between the original space and the feature space, utilizing the
structure information, e.g. covariance matrices in the feature space seems not
to do much help in this sense. Inspecting these two points, one interesting
topic in the future is to consider forcing constraints on the mapping function
so as to maintain the data topology in the kernelization process.
In the above, we do not perform the reduction on these datasets. To illustrate how the reduction algorithm works for decreasing the computation time
while maintaining the test accuracy, we implement it on the Heart-disease
dataset. We perform the reduction in training sets and then keep test sets unchanged. We repeat this process for dierent thresholds . We then plot the
curve of the cross validation accuracy against the threshold . Moreover, we
also plot the curve of the computation time against the threshold. This can
be seen in Fig. 4.8. From this gure, we can see that both that the computation time and the test accuracy change insensitively against when is set
to some small values, e.g. 0.7. If looking into the Heart-disease dataset,
we nd that most data points are far away from their corresponding class
center in terms of the Manhalanobis distance. Thus setting small values to
does not actually reduce many data points. This generates both a relatively
at changing curve in the test accuracy and the computation time in this
range. As is changing larger, the computation time decreases fast as more
and more data points are removed, while the test accuracy goes down slowly.
When the threshold is set to 1, the M4 degrades to the MPM model, yielding
the test accuracy of M4 achieves the same value of MPM. This demonstrates
how the proposed reduction algorithms can decrease the computation time
while maintaining good performance. When used in practice, the threshold
can be set according to the required response time.
(a) Test accuracy vs. threshold
(b) Running time vs. threshold
Fig. 4.8. Reduction on the Heart-disease dataset
4.7 Summary
93
4.6 Discussions and Future Work

We will discuss several important issues in this section. First, although M4
can be solved in polynomial time, the large computation time is still one of its
biggest limitations. This may cause problems especially in its kernelization
version. Note that the proposed reduction algorithm in this chapter does
not completely solve this problem, since removing points will inevitably lose
information. In this sense, it is crucial to develop some special algorithms
for M4 . Due to the sparsity of M4 (it also contains support vectors), it is
therefore very interesting to investigate whether decomposable methods or
an analogy to the Sequential Minimal Optimization [16] designed for SVM
can also be applied in training M4 . We believe that there is much to obtain
from such explorations. Certainly, this is a highly worthy research direction
in the future.
Second, although we have derived an error bound for M4 , digging out
the direct connection or performing empirical comparison of this bound with
those of its special cases is still interesting, namely, SVM and MPM maintains
an interesting problem. Especially, it is an open problem whether there exists
a unied form of the bounds for M4 , SVM, and MPM. This interesting subject
deserves future deep explorations.
Third, since in this chapter we mainly discuss M4 for two-category classications, how to extend its application to multi-way classications is also
an important topic in the future.
4.7 Summary
Local learning approaches, e.g. large margin machines have demonstrated
their advantages in machine learning and pattern recognition. However, they
derive the decision boundary only in a local way. For example, the most popular large margin classier, Support Vector Machine obtains the decision hyperplane by focusing on considering some critical local points called support
vectors, while discarding all other points; on the other hand, global learning
models (e.g. Minimax Probability Machine) obtain the classier only based
on global information, i.e. the mean and covariance information in MPM,
while ignoring all individual local points. Dierently, our proposed model
is constructed based on both domestic and global view of data. This new
model is theoretically important in the sense that SVM and MPM can both
be considered as its special cases. Furthermore, the optimization of M4 can
be cast as a sequential Conic Programming problem which can be solved in
polynomial time.
We have provided a clear geometrical interpretation, and established detailed connections among our model and other models such as Support Vector
Machine, Minimax Probability Machine, Fisher Discriminant Analysis, and
Minimum Error Minimax Probability Machine. We have also shown to exploit
94
References
Mercer kernels to extend our model to build up nonlinear decision boundaries. In addition, we have also proposed a reduction method to decrease
the computation time. Experimental results on both synthetic datasets and
real world benchmark datasets have demonstrated the advantages of M4 over
Support Vector Machine and Minimax Probability Machine.
References
1. Bertsekas DP (1999) Nonlinear Programming. Belmont, MA: Athena Scientic
2nd edition
2. Blake CL, Merz CJ (1998) Repository of machine learning databases, University
of California, Irvine, http://www.ics.uci.edu/mlearn/MLRepository.html
3. Breiman L (1997) Arcing classiers. Technical Report 460, Statistics Department, University of California
4. Fukunaga K(1990). Introduction to Statistical Pattern Recognition. San Diego,
CA: Academic Press, 2nd edition
locally and globally. In the 21st International Conference on Machine Learning
(ICML-2004)
6. Huang K, Yang H, King I, Lyu MR, Chan L (2004) The minimum error
minimax probability machine. Journal of Machine Learning Research 5:1253
1286
7. Huang K, Yang H, King I, Lyu MR, Chan L (2007). Maxi-Min Margin Machine:
Learning large margin classiers globally and locally. To appear in IEEE Trans.
Neural Networks
8. Ivannov VV (1962) On linear problems which are not well-posed. Soviet Math.
Docl. 3(4):981983
9. Kohavi R (1995). A study of cross validation and bootstrap for accuracy estimation and model selection. In Proceedings of the Fourtheenth International Joint
Conference on Articial Intelligence (IJCAI-1995). San Francisco, CA:Morgan
Kaufmann 338345
10. Kruk S, Wolkowicz H (2000) General nonlinear programming. In H. Wolkowicz,
R. Saigal, and L. Vandenberghe, editors, Handbook of Semidenite Programming: Theory, Algorithms, and Applications. Boston, MA: Kluwer Academic
Publishers 563575
3:555582
12. Lobo M, Vandenberghe L, Boyd S, Lebret H (1998) Applications of second
order cone programming. Linear Algebra and its Applications 284:193228
13. Luntz A, Brailovsky V(1969) On estimation of characters obtained in statistical
procedure of recognition (in Russian). Technicheskaya Kibernetica 3(6)
14. Marshall AW, Olkin I (1960) Multivariate Chebyshev inequalities. Annals of
15. Nesterov Y, Nemirovsky A (1994) Interior point polynomial methods in convex
programming: Theory and applications. Philadelphia, PA: SIAM
16. Platt J(1998) Sequential minimal optimization: A fast algorithm for training
support vector machines. Technical Report MSR-TR-98-14
References
95
17. Pruessner A(2003). Conic programming in GAMS. In Optimization Software

The State of the Art. INFORMS Atlanta, http://www.gamsworld.org/cone/links.htm
18. Sch
olkopf, Smola A(2002) Learning with Kernels. Cambridge, MA: The MIT
Press
19. Smola AJ , Bartlett PL , Scholkopf B, Schuurmans D(2000) Advances in Large
Margin Classiers. Cambridge, MA: The MIT Press
20. Sturm JF (1999) Using sedumi 1.02, a matlab toolbox for optimization over
21. Sturm JF (2000) Central region method. In J. B. G. Frenk, C. Roos, T. Terlaky,
and S. Zhang, editors, High Performance Optimization. Boston, MA: Kluwer
Academic Publishers 157194
22. Tikhonov AN (1963) On solving ill-posed problem and method of regularization. Doklady Akademii Nauk USSR 153:501504
23. Tong S, Koller D (2000) Restricted Bayes optimal classiers. In Proceedings of
the 17th National Conference on Articial Intelligence (AAAI), Austin, Texas
658664
24. Vapnik VN (1998) Statistical Learning Theory. New York, NY: John Wiley &
Sons
Springer Verlag, 2nd edition
26. Vasin VV (1970) Relationship of several variational methods for approximate
solutions of ill-posed problems. Math. Notes 7:161166
27. Zhu J, Rosset S, Hastie T, Tibshirani R(2003) 1-norm support vector machines.
In Advances in Neural Information Processing Systems (NIPS 16)
5
Extension I: BMPM for Imbalanced Learning
In this chapter, we consider the imbalanced learning problem. This problem

means the task of binary classication on imbalanced data, in which nearly
all the instances are labeled as one class, while far fewer instances are labeled
as the other class, usually the more important class. Traditional machine
learning methods seeking accurate performance over a full range of instances
are not suitable to deal with this problem, since they tend to classify all
the data into the majority class, usually the less important class. Moreover,
many current methods have tried to utilize some intermediate factors, e.g.
the distribution of the training set, the decision thresholds or the cost matrix,
to impose a bias towards the important class. However, it remains uncertain
whether these roundabout methods can improve the performance in a systematic way. In this chapter, we apply Biased Minimax Probability Machine,
one of the special cases of Minimum Error Minimax Probability Machine to
deal with the imbalanced learning tasks. Dierent from previous methods,
this model achieves in a worst-case scenario to derive the biased classier by
directly controlling the classication accuracy on each class. More precisely,
BMPM builds up an explicit connection between the classication accuracy
and the bias, which thus provides a rigorous treatment on imbalanced data.
We examine dierent models and compare BMPM with three other competitive methods, i.e. the Naive Bayesian classier, the k-Nearest Neighbor
method, and the decision tree method C4.5. The experimental results demonstrate the superiority of this model.
This chapter is organized as follows. In the next section, we briey present
an introduction to the imbalanced learning. We then reiterate in a tight
version the theoretical foundation of this chapter, namely the BMPM model.
Following that in Section 5.3 we apply the BMPM model to deal with the
imbalanced learning tasks. In Section 5.4, we evaluate the BMPM model
based on a series of experiments, and in Section 5.5, we make discussions and
present future work. Finally, we summarize this chapter in Section 5.6.
98
5 Extension I: BMPM for Imbalanced Learning
5.1 Introduction to Imbalanced Learning

Learning classiers from imbalanced or skewed datasets is an important topic,
arising very often in practice in classication problems. In such problems,
almost all the instances are labeled as one class, while far fewer instances
are labeled as the other class, usually the more important class. It is obvious
that traditional classiers seeking accurate performance over a full range of
instances are not suitable to deal with imbalanced learning tasks, since they
tend to classify all the data into the majority class, which is usually the less
important class.
To cope with imbalanced datasets, there are types of methods such as
the methods of sampling [4, 22, 15], the methods of moving the decision
thresholds [26, 29], and the methods of adjusting the cost matrix [3, 26].
The rst school of methods aims to reduce the data imbalance by downsampling (removing) instances from the majority class or up-sampling
(duplicating) the training instances from the minority class or both. The
second school of methods tries to adapt the decision threshold to impose a
bias on the minority class. Similarly, the third school of methods improves
the prediction performance by adjusting the weight (cost) for each class.
A common problem for all the three families of methods is that they lack
a rigorous and systematic treatment on imbalanced data. For the sampling
method, either up- or down-sampling is unsuitable: up-sampling will introduce noise, while down-sampling the data will lose information. Moreover,
to incorporate a good bias, it is usually dicult to know what a proportion
should be sampled. For these reasons, Provost stated it as an open problem
whether simply varying the skewness of the data distribution can improve
prediction performance systematically [29]. For the method of adjusting the
cost matrix or adapting weights, similar problems are also encountered, i.e.
they are hard to build direct connections between the cost matrix or the
weights and the biased classication quantitatively. To impose a suitable
bias towards the important class, they have to adapt these factors by trials.
Therefore, these methods cannot rigorously handle imbalanced data.
In this chapter, we apply Biased Minimax Probability Machine (BMPM)
to handle the tasks of learning from imbalanced data. Dierent from the sampling methods, BMPM does not remove or duplicate data. When compared
with the methods of changing the thresholds or weights, our model builds
up an explicit connection between the classication accuracy and the bias.
It thus oers an elegant way to incorporate the bias into classication by
directly controlling the real accuracy.
5.2 Biased Minimax Probability Machine

Suppose two random n-dimensional vectors x and y represent two classes of
data, where x belongs to the family of distributions with a given mean x
5.2 Biased Minimax Probability Machine
99
and a covariance x , denoted as x (

x, x ); similarly, y belongs to the
family of distributions with a given mean y
and a covariance y , denoted
as y (
y , y ). Here x, y, x
, y
Rn , and x , y Rnn . In this chapter,
the class x also represents the important or minority class and the class y
represents the corresponding less important or majority class.
The Biased Minimax Probability Machine can be described as follows1 :
max
,,b,w=0
s.t.
inf
Pr {wT x b} ,
(5.1)
inf
Pr {wT y b} ,
(5.2)
x(
x, x )
y(y,
y)
0 .
(5.3)
Here means the lower bound of the probability (accuracy) for the classication of future cases of the class x with respect to all distributions with the
mean and covariance as (x, x ); in other words, is the worst-case accuracy
for the class x. Similarly, is the lower bound of the accuracy of the class y.
This optimization achieves to maximize the accuracy (the probability ) for
the biased class x while simultaneously maintaining the class ys accuracy at
an acceptable level 0 by setting a lower bound as Eq.(5.3). In comparison,
the Minimax Probability Machine (MPM) in [16, 17] considers the balanced
dataset; therefore, it makes equal to .
This optimization setting seems to be more useful in incorporating a bias
into classications for imbalanced learning problems. A typical example can
be seen in the epidemic disease diagnosis problem which is usually an imbalanced classication problem as well. The ill cases are usually much fewer
than the healthy cases. However, misclassication of the ill class results in
more serious consequence than misclassication of the healthy case. Thus
an unequal treatment on dierent classes is obviously necessary.
We summarize the advantages of our biased model in the following. First,
this method provides a dierent treatment on dierent classes, i.e. the hyperplane w T z = b given by the solution of this optimization favors the
classication of the important class x over the less important class y. Second, given reliable mean and covariance matrices, the derived decision hyperplane is directly associated with two real accuracy indicators, i.e. and
, for each class. Thus, by varying the lower bound of , i.e. 0 and deriving
the corresponding classier, we can quantitatively incorporate a bias into the
classication. Third, this model contains a distribution-free feature. With no
distribution assumption on data, the derived hyperplane seems to be more
general and valid than a large family of classiers, namely the generative classiers [10, 12] including the Naive Bayesian classier [18], which has to make
1
Note that, for easy explanations, the model description is in the slightly dierent but essentially the same form as the one introduced in Chapter 3.
100
specic distribution assumptions. Fourth, as shown shortly in Section 5.3, either we can simply modify this BMPM optimization to automatically search
the best 0 in terms of some standard criteria, or slightly dierent from the
current setting, we can quantitatively generate the trade-o curve between
the accuracies on dierent classes and leave the task of choosing the best 0
to the users. Finally, although the BMPM contains the above advantages, it
does not trade them for eciency. It is shortly shown that the optimization of
BMPM can be cast as a Fractional Programming (FP) problem and thus can
be solved eciently. In short, with these important features, BMPM appears
to oer a more direct and rigorous scheme to handle biased classication
tasks, especially the imbalanced classications, where the importance or cost
for each class is unequal.
5.3 Learning from Imbalanced Data by Using BMPM

In this section, we apply the novel BMPM model to the tasks of learning from
imbalanced data. We rst review four standard imbalanced learning criteria,
then based on two of them, we apply BMPM to the imbalanced learning
tasks.
5.3.1 Four Criteria to Evaluate Learning from Imbalanced Data
In general, four criteria are used to evaluate the imbalanced learning. They
are (1) the criterion of Minimum Cost (MC), (2) the criterion of Maximum
Geometry Mean (MGM) of the accuracies on the majority class and the
minority class, (3) the criterion of the Maximum Sum (MS) of the accuracies
on the majority class and the minority class, and (4) the criterion of Receiver
Operating Characteristic (ROC) analysis. We review these criteria as follows.
Aiming to solve the problems caused by maximizing the accuracy over a
full range of data, instead, Grzymala-Busse, et al. [9] maximized the sum of
the accuracies on the minority class and the majority class (or maximized
the dierence between the true positive and false positive accuracy). This
criterion is also widely used in other elds, e.g. graph detection, especially line
detection and arc detection, where it is called Vector Recovery Index [6, 23].
Similarly, Kubat, et al. [14] proposed to use the geometric mean instead
of the sum of the accuracies. However, compared to maximizing the sum,
this criterion has a nonlinear form, which is not easy to be automatically
optimized. On the other hand, when the cost of misclassication is known, a
minimum cost measure dened as Eq.(5.4) should be used [2]:
Cost = Fp CFp + Fn CFn ,
(5.4)
where Fp is the number of the false positive, CFp is the cost of a false positive,
Fn is the number of the false negative, and CFn is the cost of a false negative. However, because the cost of misclassication is generally unknown in
5.3 Learning from Imbalanced Data by Using BMPM
101
real cases, the usage of this measure is somewhat restricted. Considering this
point, some researchers introduced the ROC analysis [25, 26, 34]. This criterion plots a so-called ROC curve to visualize the tradeo between the false
positive rate and the true positive rate and leaves the task of the selection
of a specic tradeo to the practitioners. Fig. 5.1 illustrates an articially
generated ROC curve. It has been suggested that the area beneath an ROC
curve can be used as a measure of accuracy in many applications [30, 33].
Thus, a good classier for imbalanced learning should have a larger area.
Fig. 5.1. An articially generated Receiver Operating Characteristic (ROC) curve
Based on the above review, in this chapter we will focus on using the
criterion of MS and the ROC curve analysis to evaluate the classiers.
5.3.2 BMPM for Maximizing the Sum of the Accuracies
In the following, we rst modify the formulation of BMPM to maximize the
sum of the accuracies for two classes. Next, we make an analysis on the
solvability of the modication version. Finally, we present the optimization
method.
5.3.2.1 Model Modication
When using BMPM for the criterion of MS, we can modify the formulation
of BMPM as follows:
max
( + ) ,
(5.5)
inf
Pr {wT x b} ,
(5.6)
inf
Pr {wT y b} .
(5.7)
,,b,w=0
s.t.
x{x, x }
y{y, y }
102
The above formulate directly maximize the sum of the lower bounds of the
accuracies so as to maximize the sum of the accuracies. In comparison, to
achieve the maximum sum of the accuracies, some other approaches, e.g. the
methods of sampling or the methods of adapting the weights have to search
the best sampling proportion or the best weights by trials, which are in
general very time-consuming. Since the above optimization is in fact nearly
the same as the Minimum Error Probability Machine, it can be similarly
solved by the Sequential Biased Minimax Probability Machine optimization
method as introduced in Chapter 3. We thus do not elaborate it here.
5.3.3 BMPM for ROC Analysis
It is straightforward to apply the BMPM model to plot the ROC curve, since
the lower bounds and directly and quantitatively control the accuracies
for two classes. We only need to adapt the acceptable level for , namely
0 , from 0 to 1, to obtain a sequence of trade-os between the accuracies
of the important class and the negative class. We address that again, since
0 represents the lower bound of the accuracy of the less important class,
varying 0 provides a direct and quantitative way to move the decision plane
with dierent trade-os. Directly associating accuracies with the moving of
the hyperplane while assuming no distribution is one of advantages of BMPM
over the other methods by adapting the weights or thresholds.
5.4 Experimental Results

In this section, we rst illustrate the BMPM model with a toy example,
and then evaluate the performance of BMPM on two real world imbalanced
datasets, namely the recidivism dataset and the rooftop dataset in comparison with the Naive Bayesian (NB) classier, the k-Nearest Neighbor (k-NN)
method [1], and the decision tree classier C4.5 [31].
5.4.1 A Toy Example
We present a toy example to illustrate the BMPM model in this section.
Suppose 15 data points of the class x are generated from a 2D Gaussian
distribution with the mean and covariance matrix as x = [0 1.5]T and x =
[0.5 0; 0 0.5] and 65 data points of the class y from another 2D Gaussian
distribution with y = [0 0]T and y = [0.5 0; 0 0.5].
By adapting the lower bound accuracy 0 for the class y, with optimizing
the corresponding BMPM, we obtain a series of decision boundaries for the
2
toy example when using the Gaussian kernel exy / with the parameter
as 5. These boundaries are illustrated in Fig. 5.2. Gray regions are classied as
the class x represented by +s, whereas those outside gray regions are judged
103
Fig. 5.2. A toy example to illustrate BMPM. Data of the class x is plotted
as +s, and data of class y as s. The gray area represents the classication
region of the class x, while the area outside the gray region is classied as
the class y
as the class y plotted as s. It is clear to observe that the lower bound 0
directly controls the accuracy of the class y. More specically, when 0 is set
to small values such as 10.00%, 60.00% and 95.00%, the boundary is biased
towards the class x. When 0 is set to larger values such as 99.00%, the
classication is biased towards the class y. Moreover, Table 5.1 demonstrates
that the lower bounds 0 and can serve as the accuracy indictors. It is
observed that these lower bounds keep well, i.e. the corresponding accuracies
are slightly higher than the lower bounders except in the case when 0 =
0.95. The exception, i.e. that the value of , 99.16% is greater than the real
accuracy 93.33%, is understandable due to the relatively smaller number of
training samples: one single misclassication will inuence the classication
results signicantly. This toy example demonstrates that by changing 0 ,
104

Table 5.1. Lower bounds of accuracies, , 0 and the real accuracies
0 (%) True negative rate(%) (%) True positive rate(%)
10.00
13.85
100.00
100.00
60.00
63.08
100.00
100.00
95.00
95.38
99.16
93.33
99.00
100.00
81.94
86.67
BMPM provides an elegant and direct way to incorporate the bias into the
classication.
5.4.2 Evaluations on Real World Imbalanced Datasets
In this section, we evaluate our novel BMPM model in comparison with three
competitive classication methods, namely the Naive Bayesian classier, the
k-Nearest Neighbor methods and the decision tree C4.5, on two real world
imbalanced datasets, the recidivism dataset and the rooftop dataset. Before
we go into the experimental details, we rst introduce these three techniques
and adapt them to learn from imbalanced datasets according to previous
research results [20, 26].
5.4.2.1 Modifying Three Learning Techniques
We investigate and modify three learning techniques, the Naive Bayesian
classier, the k-Nearest Neighbor method, and the decision tree C4.5 in the
following.
The Naive Bayesian classier [11, 18] is proposed based on a very simple assumption, i.e. each attribute is conditionally independent of each
other when given the class variable. The decision in a two-category prediction task is made according to the calculation of the posterior probability
p(C|z), where C is the class variable and z represents the observation. When
p(C1 |z) 0.5 or another equivalent yet more convenient rule is satised,
i.e. p(C1 )p(z|C1 ) p(C2 )p(z|C2 ), z is classied into C1 ; otherwise, it is
judged as C2 . Even with the strong conditional independency assumption,
the Naive Bayesian classier demonstrates a surprisingly good performance
when compared with state-of-the-art classiers [8, 19] such as Support Vector
Machines [35] and C4.5 in many domains. By simply introducing a parameter
into the decision rule p(C1 )p(z|C1 ) p(C2 )p(z|C2 ), Naive Bayesian classiers can be adapted to the imbalanced learning. For example, specifying
< 1 imposes a bias towards the C1 class, whereas specifying > 1 imposes
a bias towards the C2 class.
In the k-Nearest Neighbor classication [1], based on some distance measure, e.g. the Euclidean distance measure, k data points, which are the closest to the query point, are selected out. It then labels the query point as
105
the most frequent class among the chosen k points. Although this method is
very simple and may suer from diculties in high dimensions, it achieves
satisfactory performance in many real domains. Following [26], we alter the
distance measure j for the class Cj to handle imbalanced learning tasks
according to Eq.(5.8):
j = dE (z, z j ) j dE (z, z j ) ,
(5.8)
where z j is the closest point from class Cj to the query point, and dE (z, z j )
represents the Euclidean distance measure. Similar to the Naive Bayesian
classier, by modifying j the Nearest Neighbor method can build biased
classiers.
C4.5 is a kind of algorithm introduced by Quinlan for inducing classication models, also called decision trees, from data [31]. By selecting the
attributes according to the gain ratios criterion, an information measure of
homogeneity, C4.5 builds up a decision tree where each path from the root
to a leaf represents a specic classication rule. We adapt C4.5 to learn from
imbalanced dataset based on the similar method to [26], i.e. by changing the
prior probability to bias the classication.
5.4.2.2 Evaluations on the Recidivism Dataset
The recidivism dataset was obtained from a cohort of releases of the North
Carolina prison system during the time period from July 1, 1977 to June
30, 1978. There are totally 4, 618 individuals in this dataset, including a
training set with 1, 540 individuals and a test set with 3, 078 individuals. In
the training set, 570 (27.5%) individuals were recidivists and 970 (72.5%) were
not. In the test set, 1, 151 individuals were recidivists and 1, 927 were not.
Although this dataset is not skewed as severely as other reported datasets,
for example, the fog dataset [28] and the rooftop dataset used in the next
subsection, it is enough to use this dataset to evaluate the performance of
the imbalanced learning [26].
We use the same processing method [32] to select and scale nine attributes
that appear in Table 5.2, while six other attributes are dropped based on an
insignicant test at the 5% level.
We compare the performance of our proposed Biased Minimax Probability Machine model, in both the linear (BMPML) and the Gaussian kernel
setting (BMPMG), with the Naive Bayesian classier, C4.5 and the k-Nearest
Neighbor method. These methods are modied into the imbalanced learning
according to the methods introduced in the previous section. We run k-NN
methods for k = 1, 3, 5, . . . , 21, but we only present the best three results
for brevity. The width parameter for the Gaussian kernel is tuned via cross
validation methods [13].
We rst present the experimental results based on the MS criterion in
Table 5.3. To be more comparable, we show the average of the accuracy for
106

Table 5.2. Attribute description in the recidivism dataset
Attribute
Description
TSERVED
Time served (in months)
AGE
Age (in months) at the time of release
PRIORS
Number of previous incarcerations
WHITE
Is the individual Caucasian?
FELON
Was the sentence for a felony?
LCHY
Does individuals record indicate a serious problem with alcohol?
JUNKY
Does individuals record indicate a serious problem with hard drugs?
PROPTY
Was individuals sentence for a crime against property?
MALE
Is the individual male?
each class when each classier attains the point of the maximum sum. The
BMPML achieves an average accuracy of 0.6391 and the BMPMG achieves an
average accuracy of 0.6490, while the highest average accuracy among other
classiers is given as 0.6272 by NB. Therefore, in this dataset, BMPML and
BMPMG outperform other methods in terms of the MS criterion.
Table 5.3. Performance on a recidivism prediction task based on the MS
criterion
Method True negative rate True positive rate (True positive rate+true negative rate)/2
NB
0.6177
0.6377
0.6272
k-NN(9)
0.6255
0.5464
0.5860
k-NN(11)
0.6238
0.5542
0.5890
k-NN(13)
0.5569
0.6201
0.5885
C4.5
0.7405
0.4900
0.6153
BMPML
0.7037
0.5745
0.6391
BMPMG
0.7203
0.5778
0.6490
Let us next present the experimental results based on the ROC analysis. By setting the thresholds or costs by trials for NB, k-NN, and C4.5, the
ROC curves are generated with good shapes as evenly distributed along their
length as possible. As discussed in [26], although this generation method may
increase the running time for some methods, e.g. k-NN, it works well in C4.5
and NB and is sucient to evaluate the performance of imbalanced learning.
For the BMPM model, since the lower bound 0 serves as the accuracy indicators, we simply vary it from 0 to 1 to generate the corresponding ROC
curve. The ROC curves are shown in Fig. 5.3(a). As seen in this gure, the
performances of BMPML and BMPMG are once again superior to those of
107
Fig. 5.3. ROC curves for the recidivism dataset. Subgure (a)
shows a full range of the ROC curve, while (b) shows a critical
proportion of the ROC curve, which is of more interest in real applications. Both gures demonstrate the superiority of the BMPM
model, since the curves of BMPML and BMPMG cover those of
other models in most parts and thus have a larger area
other methods, since their ROC curves cover those of other models in most
parts. To quantitatively demonstrate the dierence, in Table 5.4 we also show
the areas beneath the ROC curves approximated by using the trapezoid rule.
The BMPML and BMPMG show a consistent superiority to NB which is the
best of the other three methods.
In addition, in real applications not all the portions of the ROC curve are
of great interest [27]. Usually, those with a small false positive rate and a high
true positive rate should be more of interest and importance [36]. We thus
108

Table 5.4. Performance on a recidivism prediction task
based on the area of ROC curve
Method
Area under ROC curve
NB
0.6646
k-NN(11)
0.6155
k-NN(13)
0.6189
k-NN(17)
0.6148
C4.5
0.6383
BMPML
0.6842
BMPMG
0.6798
especially show the portion of the ROC curve in the range when the false
positive rate FP [0, 0.5] and the true positive rate TP [0.5, 1]. As shown
in Fig. 5.3(b), in this range, the superiority of the BMPL and BMPMG is
more obvious than the whole ROC curve analysis. This again demonstrates
our models advantages over other methods.
5.4.2.3 Evaluations on the Rooftop Dataset
The rooftop dataset consists of 17, 829 overhead images of Fort Hood, Texas,
collected as part of the RADIUS project [7], which are of a military base.
Depending on whether they are buildings (with a detected rooftop) or not,
781 images in this dataset are labeled as positive examples while 17, 048
images are labeled as negative examples. It is clearly observed that this is
a severely skewed dataset. According to [7, 26], these images were taken
from two dierent viewpoints, i.e. a nadir aspect and an oblique aspect and
covered three dierent areas. Following [21, 26], we represent each of these
images in nine continuous attributes which are extracted based on various
image analysis. The detailed information about this dataset is summarized
in Tables 5.5 and 5.6.
Table 5.5. Description of images in the rooftop dataset
Sub-dataset
Location
Aspect
#Positive
#Negative
Image size
2055 375
Nadir
71
2645
1803 429
Oblique
74
3349
670 645
Nadir
197
982
704 568
Oblique
238
1955
1322 642
Nadir
87
3722
1534 705
Oblique
114
4395
109
Table 5.6. Description of the attributes in the rooftop dataset

Attribute
Description
Evaluation of the edge support
Evaluation of the corner support
Evaluation of the parallel support
Evaluation of the OTV (Orthogonal Trihedral Vertex) support
Evaluation of the shadow corner support
Evaluation of gap overlap
Evaluation of displacement of edge support
Evaluation of crossing lines on any side of the hypothesis
Evaluation of existence of T-junction or L-junction on any side
We randomly split the rooftop data into a training set with 60% data and
a test set with 40% data. We then construct classiers from imbalanced data
based on the training dataset and perform evaluations on the test dataset.
We repeat this procedure ten times and use the average of the results as the
performance metric. In such a setup, we compare our BMPM with other three
approaches, i.e. NB, C4.5 and k-NN. Similar to the case in the recidivism
dataset, NB, C4.5 and k-NN are modied to handle imbalanced data. The
width parameter is chosen by cross validation methods again. Moreover, we
still run k-NN with k = 1, 3, 5, ..., 21 and present the best three for brevity.
The results are summarized in Table 5.7 based on the MS criterion, and
Table 5.7. Performance on the rooftop dataset based on the MS criterion
Method True negative rate
True positive rate (True positive rate + True negative rate)/2
BMPML
0.8015 0.0058
0.8231 0.0063
0.8123 0.0060
BMPMG
0.7997 0.0087
0.8405 0.0100
0.8201 0.0091
k-NN(7)
0.7510 0.0055
0.8069 0.0062
0.7789 0.0052
k-NN(13)
0.7409 0.0051
0.8140 0.0083
0.7774 0.0061
k-NN(15)
0.7433 0.0067
0.8211 0.0072
0.7822 0.0072
NB
0.7969 0.0043
0.8177 0.0080
0.8073 0.0066
C4.5
0.8176 0.0040
0.7942 0.0063
0.8059 0.0051
Fig. 5.4 and Table 5.8 based on the ROC analysis. As is clearly observed, for
both criteria, the BMPM method demonstrates its superiority to the other
methods, since it has higher sums of the accuracies and larger areas under the
ROC curves. Similar to what we do in the recivisim dataset, we also plot the
more critical portion of the ROC curve in Fig. 5.4(b). The predominance of
BMPML and the BMPMG is even more obvious. To evaluate the performance
more reliably, we perform a signicance test based on both LabMRMC [5, 24]
110
and a t-test. The analysis shows that the accuracies of BMPML and BMPMG
are signicantly dierent from those of other methods at P 0.05, both in
terms of the MS criterion and the ROC curve criterion.
Fig. 5.4. ROC curves for the rooftop dataset. We ran each method by
randomly partitioning the dataset into a training dataset (60%) and a test
dataset (40%). The evaluations were iterated 10 times. We then average
the true positive rate and false positive rate to generate the ROC curves.
Subgure (a) shows a full range of the ROC curve, while (b) shows a critical
proportion of the ROC curve, which is of more interest in real applications.
Both gures demonstrate the superiority of the BMPML and BMPMG
model to other models, since the curves of BMPML and BMPMG cover
those of other models in most parts and thus have a larger area
111
Table 5.8. Performance on the rooftop dataset based on

the area of ROC curve
Method
Area under ROC curve
BMPML
0.8791 0.0061
BMPMG
0.8819 0.0087
k-NN(9)
0.8601 0.0091
k-NN(11)
0.8569 0.0058
kNN(15)
0.8582 0.0063
NB
0.8678 0.0060
C4.5
0.8744 0.0062
5.4.3 Evaluations on Disease Datasets

Diagnosing diseases contain a very similar characteristic to the imbalanced
learning, since one class, usually the disease class needs to be given more bias
than the other class. Therefore, the above discussed model modications will
be automatically applicable for this kind of tasks. In the following, we evaluate the performance of BMPM on two disease datasets, namely, the Breastcancer dataset and the Heart-disease dataset, which are obtained from UCI
machine learning repository. In the context of diagnosing diseases, the true
positive rate is usually called sensitivity, while the true negative rate is called
specicity. Therefore, we should maximize the sensitivity while maintaining
the specicity acceptable. In the following, we present the experimental results still compared with the best three, namely the modied Naive Bayesian
classier, k-NN, and C4.5. We randomly split the data for each dataset into a
training set with 80% data and a test set with 20% data. We then construct
classiers based on the training dataset and perform evaluations on the test
dataset. We repeat this procedure ten times and use the average of the results
as the performance metric.
We present the results based on the MS criterion in Table 5.9 for the
breast-cancer dataset and Table 5.10 for the heart disease dataset. Obsereved
from these two tables, the BMPM model also demonstrates a superiority to
other three models. In addition, the t-test also shows that the accuracies of
BMPML and BMPMG are signicantly dierent from those of other three
classiers at P 0.05.
We next present the experimental results based on the ROC analysis
in Fig. 5.5(a) and Fig. 5.6(a). It is observed that BMPML and BMPMG
perform better than other classiers for both datasets, since in most parts
the BMPM curves dominate those of other methods. More specically, we
calculate the areas under the ROC curves as illustrated in Table 5.11, based
on the trapezoid rule. For the breast-cancer dataset, it produces a curve with
an area of 0.9953 in the linear setting and a curve with an area of 0.9963 in
112

Table 5.9. Comparison of the model performance based on the
MS criterion on the breast-cancer dataset
Method
Specicity
Sensitivity
(Specicity+Sensitivity)/2
BMPML
0.9684 0.0029
0.9872 0.0015
0.9778 0.0021
BMPMG
0.9612 0.0018
0.9915 0.0011
0.9764 0.0016
k-NN(11)
0.9900 0.0047
0.9620 0.0034
0.9760 0.0029
k-NN(17)
0.9862 0.0081
0.9664 0.0058
0.9762 0.0050
k-NN(7)
0.9721 0.0071
0.9752 0.0049
0.9737 0.0058
NB
0.9366 0.0059
0.9719 0.0049
0.9543 0.0051
C4.5
0.9378 0.0074
0.9582 0.0067
0.9480 0.0072
Table 5.10. Comparison of the model performance based on the

MS criterion on the heart disease dataset
Method
Specicity
Sensitivity
(Specicity+Sensitivity)/2
BMPML
0.8549 0.0042
0.8158 0.0013
0.8354 0.0035
BMPMG
0.8403 0.0053
0.8572 0.0017
0.8488 0.0026
k-NN(17)
0.7654 0.0029
0.8837 0.0018
0.8246 0.0027
k-NN(7)
0.7754 0.0038
0.8844 0.0042
0.8299 0.0037
k-NN(15)
0.7512 0.0028
0.8653 0.0037
0.8082 0.0036
NB
0.7862 0.0052
0.8024 0.0031
0.7943 0.0040
C4.5
0.8831 0.0022
0.7065 0.0018
0.7948 0.0021
the Gaussian kernel, whereas the k-NN with k = 11 forms a curve with a
smaller area equal to 0.9908, the best result of the k-NN, NB and C4.5. For
the Heart disease dataset, the BMPM shows a curve with an area of 0.8814
in the linear setting and a curve with an area of 0.8932 in the Gaussian kernel
setting. These two areas are both greater than those of the other methods,
i.e. the k-NN classier, NB and C4.5. In summary, the evaluations based on
the area of the ROC curve quantitatively demonstrate the superiority of our
BMPM model for both datasets.
In addition, as illustrated in Fig. 5.5(b) and Fig. 5.6(b), we show the
critical portion of Fig. 5.5(a) and Fig. 5.6(a) respectively when the false
positive rate is in the range of 0.0 to 0.5 and the true positive rate is in
the range of 0.5 to 1.0. In this critical region, most parts of the ROC curves
of BMPM cover the corresponding curves of other models in both datasets,
which again demonstrates the superiority of the BMPM model.

Table 5.11. Comparison of the model performance based
on the ROC analysis
Method
Area under ROC Curve

Breast-cancer
Heart
0.9953 0.0018
0.8814 0.0056
BMPMG 0.9963 0.0016
0.8932 0.0043
BMPML
k-NN(11) 0.9908 0.0060 0.8701 0.0038

k-NN(17) 0.9902 0.0100
0.8689 0.0050
k-NN(7)
0.9887 0.0080 0.8596 0.0038
NB
0.9841 0.0060 0.8162 0.0034
C4.5
0.9762 0.0120
0.8301 0.0038
Fig. 5.5. ROC curves for the breast-cancer dataset. The ROC
curves of BMPML and BMPMG dominate those of other models
and BMPMG yields the largest area under the ROC curve
113
114
Fig. 5.6. ROC curves for the heart disease dataset. The ROC
curves of BMPML and BMPMG dominate those of other models
and BMPMG yields the largest area under the ROC curve
5.5 When the Cost for Each Class Is Known

There exists cases in which the cost for each class can be given by experts.
In the following, we show that the BMPM model can naturally be adapted
to this type of tasks.
Assuming x and y are the minority class and the majority class respectively, it is easily veried that minimizing the optimization function given by
Eq.(5.4) is equivalent to maximizing the following formulation:
max (rx Kx + ry Ky ) ,
where rx is the true positive rate or the accuracy of the class x, ry is the true
negative rate or the accuracy of the class y, Kx and Ky are two constants
which are equal to CFp Ny and CFn Nx respectively (Nx , Ny are respectively
the number of data points labeled as the classes x and y). Similar to the
References
115
optimization procedure of MS, we can naturally modify the BMPM model in

the following formulation:
max
,,b,w=0
s.t.
(Kx + Ky ) ,
inf
Pr {wT x b} ,
inf
Pr {wT y b} .
x{x, x }
y{y, y }
The above optimization derives the classication boundary by maximizing

the weighted lower bound of the real accuracies or the weighted worst-case
real accuracies so as to minimize the overall classication risk. Moreover,
similar to the MS case, it is easily validated that this optimization problem
can be cast as a sequential BMPM problem. Hence, it can similarly be solved
based on the method presented in Chapter 3.
5.6 Summary
In this chapter, we have applied a novel model named Biased Minimax Probability Machine to deal with the task of learning from imbalanced datasets.
Given reliable estimation of the mean and covariance of data, this model constructs the classication boundary by directly controlling the lower bound of
the real accuracy and thus provides a systematic and rigorous treatment
on skewed data. We have evaluated the BMPM model on two real world
imbalanced datasets and two disease datasets in terms of two criteria. In
both criteria, the performances are shown to be the best when compared
with other competitive methods such as the Naive Bayesian classier, the
k-Nearest Neighbor method, and the decision tree classier, C4.5.
References
1. Aha D, Kibler D, Albert M (1991) Instance-based learning algorithms. Machine
Learning 6:3766
2. Bradley A (1997) The use of the area under the ROC curve in the evaluation
of machine learning algorithm. Pattern Recognition 30(7):11451159
3. Cardie C, Howe N (1997) Improving minority class prediction using case specic
feature weights. In Proceedings of the Fourteenth International Conference on
Machine Learning (ICML-1997). San Francisco, CA: Morgan Kaufmann 5765
4. Chawla N, Bowyer K, Hall L, Kegelmeyer W (2002) Smote: synthetic minority
over-sampling technique. Journal of Articial Intelligence Research 16:321357
5. Dorfman K, Berbaum D, Metz C (1992) Receiver operating characteristic
rating analysis: generalization to the population of readers and patients with
the jackknife method. Investigative Radiology 27:723731
116
References
6. Dori D, Liu W (1999) Sparse pixel vectorization: An algorithm and its performance evaluation. IEEE Trans. Pattern Analysis and Machine Intelligence
21:202215
7. Firschein O, Strat T (1996) RADIUS: Image understanding for imagery intelligence. San Francisco, CA: Morgan Kaufmann
9. Grzymala-Busse JW, Goodwin LK, Zhang X (2003) Increasing sensitivity of
preterm birth by changing rule strengths. Pattern Recognition Letters 24:903
910
10. Huang K, King I, Lyu MR (2003) Discriminative training of Bayesian chow-liu
tree multinet classiers. In Proceedings of International Joint Conference on
Neural Network (IJCNN-2003), Oregon, Portland, U.S.A. 1:484488
11. Huang K, King I, Lyu MR (2003) Finite mixture model of bound semi-naive
Bayesian network classier. In Proceedings of the International Conference on
Articial Neural Networks (ICANN-2003), Lecture Notes in Articial Intelligence, Long paper. Heidelberg: Springer-Verlag 2714:115122
12. Jaakkola TS, Haussler D (1998) Exploiting generative models in discriminative
classiers. In Advances in Neural Information Processing Systems (NIPS)
13. Kohavi R (1995) A study of cross validation and bootstrap for accuracy estimation and model selection. In Proceedings of the Fourteenth International
Joint Conference on Articial Intelligence (IJCAI-1995). San Francisco, CA:
Morgan Kaufmann 338345
14. Kubat M, Holte R, Matwin S (1998) Machine learning for the detection of oil
spills in satellite radar images. Machine Learning 30(2-3):195215
15. Kubat M, Matwin S (1997) Addressing the curse of imbalanced training sets:
One-sided selection. In Proceedings of the Fourteenth International Conference
on Machine Learning (ICML-1997). San Francisco, CA: Morgan Kaufmann
179186
16. Lanckriet GRG, Ghaoui LE, Bhattacharyya C, Jordan MI (2001) Minimax
(NIPS)
3:555582
18. Langley P, Iba W, Thompson K (1992) An analysis of Bayesian classiers. In
Proceedings of National Conference on Articial Intelligence 223228
19. Lerner B, Lawrence ND (2001) A comparison of state-of-the-art classication
techniques with application to cytogenetics. Neural Computing and Applications 10(1):3947
20. Lewis D, Catlett J (1994) Heterogeneous uncertainty sampling for supervised
learning. In Proceedings of the Eleventh International Conference on Machine
Learning (ICML-1994). San Francisco, CA: Morgan Kaufmann 148156
21. Lin C, Nevatia R (1998) Building detection and description from a single
intensity image. Computer Vision and Image Understanding 72:101121
22. Ling C, Li C (1998) Data mining for direct marketing:problems and solutions.
In Proceedings of the Fourth International Conference on Knowledge Discovery
and Data Mining (KDD-1998). Menlo Park, CA: AAAI Press 7379
23. Liu W, Dori D (1997) A protocol for performance evaluation of line detection
algorithms. Machine Vision and Application 9:240250
References
117
24. Maloof MA (2002) On machine learning, ROC analysis, statistical tests of signicance. In Proceedings of the Sixteenth International Conference on Pattern
Recognition. Los Alamitos, CA: IEEE Press 204207
25. Maloof MA (2003) Learning when data sets are imbanlanced and when costs are
unequal and unknown. In Proceedings of International Conference on Machine
Learning (ICML-2003)
26. Maloof MA, Langley P, Binford TO, Nevatia R, Sage S (2003) Improved rooftop
detection in aerial images with machine learning. Machine Learning 53:157191
27. Mcclish D (1989) Analyzing a portion of the ROC curve. Medical Decision
Making 9(3):190195
28. Nugroho AS, Kuroyanagi S, Iwata A (2002) A solution for imbalanced training
sets problem by combnet and its application on fog forecasting. IEICE TRANS.
INF. & SYST, E85-D(7)
29. Provost F (2000) Learning from imbanlanced data sets. In Proceedings of the
Seventeenth National Conference on Articial Intelligence (AAAI 2000)
30. Provost F, Fawcett T (1997) Analysis and visulization of classier performance:
comparison under imprecise class and cost distributions. In Proceedings of
the Third International Conference on Knowledge Discovery and Data Mining.
Menlo Park, CA: AAAI Press 4348
31. Quinlan JR (1993) C4.5: Programs for Machine Learning. San Mateo, CA:
Morgan Kaufmann Publishers
32. Schmidt P, Witte A (1988) Predicting Recidivism Using Survival Models. New
York, NY: Spring-Verlag
33. Swets J (1988) Measureing the accuracy of diagnostic systems. Science
240:12851293
34. Swets J, Pickett R (1982) Evaluation of Diagnoistic Systems: Methods from
Signal Detection Theory. New York, NY: Springer-Verlag
Springer-Verlag, 2nd edition
36. Woods K, Kegelmeyer Jr WP, Bowyer K (1997) Combination of multiple classiers using local accuracy estimates. IEEE Tansactions on Pattern Analysis
and Machine Intelligence 19(4):405410
6
Extension II: A Regression Model from M4
In this chapter, we present a novel regression model which is directly motivated from the Maxi-Min Margin Machine(M4 ) model described in Chapter 4.
Regression is one of the problems in supervised learning. The objective is to
learn a model from a given dataset, {(x1 , y1 ), . . . , (xN , yN )}, and then based
on the learned model, to make accurate predictions of y for future values of x.
Support Vector Regression (SVR), a successful method in dealing with this
problem contains the good generalization ability [20, 17, 8, 6]. The standard
SVR adopts the 2 -norm to control the functional complexity and chooses an
-insensitive loss function with a xed tube (margin) to measure the empirical risk. By introducing the 2 -norm, the optimization problem in SVR can
be transformed to a quadratic programming problem. On the other hand, the
-tube has the ability to tolerate noise in data and xing the tube enjoys the
advantages of simplicity. These settings are in a global fashion and are eective in common applications, but they lack the ability and the exibility to
capture the local trend in some applications. For example, in stock markets,
the data are highly volatile and the associated variance of noise varies over
time. In such cases, xing the tube cannot capture the local trend of data
and cannot tolerate the noise adaptively.
One typical illustration can be seen in Fig. 6.1. In this gure, the data
contain larger noise as the x value of the data becomes larger. However, the
SVR cannot exibly and suitably handle it. As shown in Fig. 6.1(a), with a
xed -margin (set to 0.04) SVR considers the data globally and equally: The
derived approximating function in SVR deviates from the actual data trend.
On the other hand, as illustrated in Fig. 6.1(b), if we adequately consider
the local volatility of data by adaptively and automatically setting a small
margin in low volatile regions and a larger margin in high volatile areas, the
resulting approximating function (the solid line in Fig. 6.1(b)) would be more
suitable and reasonable.
Targeting to solve these problems, we propose the Local Support Vector
Regression (LSVR) model. We will show that with consideration of the local
120
6 Extension II: A Regression Model from M4
Fig. 6.1. Illustration of the -insensitive loss function with xed and nonxed margins in the feature space. In (b), a non-xed margin setting is more
reasonable. It can moderate the eect of the noise by enlarging (shrinking)
the margin width in the local area with large (small) variance of noise
data trend, our model provides a systematic and automatic scheme to locally
and exibly adapt the margin. Moreover, we will also demonstrate that this
novel LSVR model can derive special cases, containing a very similar physical
meaning to the standard SVR. Another critical feature of our model is that
the associated optimization of LSVR can be cast as a Second Order Cone
Programming (SOCP) problem which can be eciently solved in polynomial
time [11]. The margin setting in the novel LSVR model is dierent from that
in our previous work [21]. Concretely, the tube here is adapted directly based
on the functional complexity and the local trend of data. This hence provides
a more systematic and more rigorous way to moderate the margin automatically. This model can be seen as an extension to the regression model of
M4 . In M4 , the main purpose is to build a classication boundary for dierent classes, while in LSVR the goal is to model a function approximating the
data. Therefore, M4 considers dierent data trends for dierent classes, while
LSVR focuses on employing dierent data trends in dierent data regions.
This is more valuable with the framework of regression tasks.
The rest of this chapter is organized as follows: the linear LSVR model
with its theoretical background is presented in Section 6.1. In Section 6.2, we
demonstrate how the standard SVR can be considered as the special case of
our proposed model. In Section 6.3, we show the link between our proposed
LSVR model and the general large margin classier M4 . The kernelized LSVR
is tackled by utilizing the Mercers kernel in Section 6.5. Section 6.6 provides
an additional interpretation on the issue of controlling the complexity of the
LSVR model. Section 6.7 presents the experiments on both synthetic and
real data. The chapter is concluded in Section 6.8.
6.1 A Local Support Vector Regression Model
121
6.1 A Local Support Vector Regression Model

In this section, we rst present the problem and model denition of the LSVR
model. We then detail its interpretation and its appealing characteristics.
After that we state its corresponding optimization method.
6.1.1 Problem and Model Denition
A basic idea to avoid overtting in function approximation is to restrict the
class of admissible solutions by a regularization term. A common method
is to nd a function, f : Rd R, based on an N -instance dataset D =
{(xi , yi ) | xi Rd , yi R, i = 1, . . . , N } by minimizing the following
regularized functional risk:
Rreg [f ] = [f ] + C Remp [f ],
where C > 0 is a regularization parameter used as the tradeo between the
minimal empirical risk Remp [f ] and the smoothness or functional complexity
controlled by [f ].
Support Vector Regression is a successful regression model following this
idea. It attempts to nd an approximating function in the linear form:
f (x) = wT x + b,
w, x Rd , b R.
(6.1)
For the complexity term [f ], SVR selects 2 -norm or other p -norm of w. To

measure the empirical risk Remp [f ], the standard SVR uses an -insensitive
loss function [20].
In order to improve the exibility of the standard SVR, we propose a
new regression model, namely Local Support Vector Regression (LSVR). The
objective is to learn the function in Eq.(6.1) approximating the data in D
by making the function locally as less volatile as possible while keeping the
error as small as possible. We formulate this objective as follows:
N
N

1
T
w
w
+
C
(i + i ) ,
i
w,b,i ,i N
i=1
i=1
T
s.t. yi (w xi + b) wT i w + i ,
(wT xi + b) yi wT i w + i ,
i 0, i 0, i = 1, . . . , N,
min
(6.2)
(6.3)
where i and i are the corresponding up-side and down-side errors at the
i-th point, respectively, is a positive constant, i is the covariance matrix
formed by the i-th data point and those data points close to it.
122
6.1.2 Interpretations and Appealing Properties

In this section, beginning with stating the physical meaning of the term,
wT i w, we interpret our novel LSVR model.
Suppose yi = wT xi + b and yi = wT xi + b. We have the variance around
the i-th data point as
i =
k
k

1
1
(yi+j yi )2 =
[wT (xi+j xi )]2 = wT i w,
2k + 1
2k + 1
j=k
j=k
where 2k is the number of data points closest to the i-th data point. Therefore, i = wT i w actually captures the volatility in the local region around
the i-th data point. In addition, i can also measure the local functional
complexity around the i-th data, since it reects the smoothness of the corresponding local region. This will be in details addressed later in Section 6.6.
By using the rst meaning of i = wT i w (representing the local volatility), LSVR can systematically and automatically vary the tube: If the i-th
data point
lies in the area with a larger variance of noise, it will contribute to
a larger wT i w or a larger local margin. This will result in reducing the
impact of the noise around the point; on the other hand, in the case that the
i-th data point
is in the region with a smaller variance of noise, the local margin (tube), wT i w, will be smaller. Therefore, the corresponding point
would contribute more in the tting process. In comparison, the standard
SVR adopts a xed margin, which treats each point equally and therefore
lacks the ability to tolerate the change in noise.
By engaging the second compelling property of i = wT i w, namely,
a measure in describing the local functional complexity, LSVR controls the
overall smoothness of the approximating function by minimizing the average
of i as seen in Eq.(6.2). Intuitively, the margin around each point can be
neither too large nor too small: If the margin is too large, the local data
trend may not be captured for over-tolerating data; if the margin is too
small, the local data trend may be over-emphasized resulting in a highly
zig-zag approximating curve. Therefore by adding the regularization term, a
trade-o can be achieved via adapting the parameter C.
6.2 Connection with Support Vector Regression

We now analyze the connection of the LSVR model with the standard Support Vector Regression model. By considering the data trend globally and
equally, i.e. setting i = , for i = 1, . . . , N , we can transform the optimization of Eq.(6.2) as follows:
6.2 Connection with Support Vector Regression
min
w,b,i ,i
wT w + C
N

123
(i + i ) ,
i=1
s.t. yi (w xi + b) wT w + i ,
(wT xi + b) yi wT w + i ,
i 0, i 0, i = 1, . . . , N .
T
Further, if = I, we obtain:

min
w,b,i ,i
w + C
N

(6.4)

(i +
i )
(6.5)
i=1
s.t. yi (wxi + b) w + i ,
(wxi + b) yi w + i ,
i 0, i 0, i = 1, . . . , N .
(6.6)
The above optimization problem is very similar to the 1 -norm SVR, except
that it has a margin related to the complexity term. In the following, we will
prove that the above optimization is actually equivalent to the 1 -norm SVR
in a meaningful sense.
Lemma 6.1. The LSVR model with setting i = I is equivalent to the 1 norm SVR in the sense that: (1) Assuming a unique 1 exists for making 1 norm SVR optimal (i.e. setting to 1 will make the objective function minimal), if for 1 the 1 -norm SVR achieves a solution {w , b } = SVR(1 ), then

the LSVR can produce the same solution by setting the parameter = w1 ,

i.e. LSVR( w1 ) = SVR(1 ); (2) Assuming a unique 2 exists for making

1
the special case of LSVR optimal (i.e. setting to 2 will make the objective function minimal), if for 2 the special case of LSVR achieves a solution
{w2 , b2 } = LSVR(2 ), then the 1 -norm SVR can produce the same solution
by setting the parameter = 2 w2 , i.e. SVR(2 w2 ) = LSVR(1 ).
Proof. Since (1) and (2) are very similar statements, we only prove (1).

When of the special case of LSVR is setting to w1 , the value of the objec1
tive function of LSVR will be at least smaller than the one by simply setting
{w, b} = {w1 , b1 }, since {w1 , b1 } is easily veried to satisfy the constraints
of LSVR. Namely,

1
LSVR
SVR(1 ) ,
(6.7)
w1
where we use to represent superior to. We assume the solution for =

w2
1
in LSVR as {w2 , b2 }. Similarly, by setting = 1 w
in SVR, we have:
w
1
1

1
w2
LSVR
.
(6.8)
SVR 1
w1
w1
124
Combining Eqs.(6.7) and (6.8), we have:

1
w 2
LSVR
SVR (1 ) .
SVR 1
w1
w1
(6.9)
Since 1 is the unique making the objective of SVR minimal, Eq.(6.9) implies
that w2 = w1 .
In addition, if in LSVR we use the item of wT w instead of its square
root form as the structure risk or complexity risk, a similar proof will also be
applicable that the 2 -norm SVR is equivalent to the special case of LSVR
with i = . In summary, we can see that the LSVR model actually contains
the standard SVR model as special cases.
6.3 Link with Maxi-Min Margin Machine

The LSVR model can also be considered as an extension of the general large
margin classier, Maxi-Min Margin Machine (M4 ) presented previously in
this book or [10]. Within the framework of binary classications for class x
and y, the M4 is formulated as follows:
max
,w=0,b
(wT xi + b)
, i = 1, 2, . . . , Nx ,
s.t.
wT x w
(wT y j + b)
, j = 1, 2, . . . , Ny ,
wT y w
(6.10)
(6.11)
(6.12)
where x and y refer to the covariance matrices of the x and the y data,
respectively.
Within the framework of classications, M4 considers dierent data trends
for dierent classes. Analogously, in the novel LSVR model we allow dierent
data trends for dierent regions, which is more suitable for the regression
purpose.
6.4 Optimization Method

In order to solve the optimization problem of Eq.(6.2), we introduce auxiliary
variables, t1 ,. . . , tN , and transform the problem as follows:
6.5 Kernelization
125

N
N

1
min
ti + C
(i + i ) ,
w,b,ti ,i ,i
N i=1
i=1
s.t. yi (wT xi + b) wT i w + i ,
(wT xi + b) yi wT i w + i ,
w T i w ti ,
ti 0, i 0, i 0, i = 1, . . . , N .
(6.13)
(6.14)
(6.15)
It is clear that Eqs.(6.14) and (6.15) are non-convex constraints. This

may present diculties in optimizing the LSVR problems. In the following,
we relax the optimization
to a Second Order Cone Programming (SOCP)

problem [11] by replacing wT i w with its upper bound ti :

N
N

1
ti + C
(i + i ) ,
min
w,b,ti ,i ,i
N i=1
i=1
s.t. yi (wT xi + b) ti + i ,
(wT xi + b) yi ti + i ,
w T i w ti ,
ti 0, i 0, i 0, i = 1, . . . , N .
Since ti is closely related to wT i w, weighting the margin width with

ti will contain a meaning similar to the original motivation, i.e. adapting
the margin exibly. More importantly, the relaxed form is a linear programming problem under quadratic cone constraints, or more specically it is a
Second Order Cone Programming. Therefore, this problem can be solved in
polynomial time by many general optimization packages, e.g. Sedumi [18, 19].
6.5 Kernelization
In this section we extend the above linear regression model to the non-linear
one by using the Mercers kernel. Suppose the training data are mapped into
a kernel space or a feature space by the mapping function, : Rd Rf .
Then, the objective in the feature space is transformed as follows:

N
N

1
min
ti + C
(i + i ) ,
(6.16)
w,b,ti ,i ,i
N i=1
i=1
s.t. yi (wT (xi ) + b) ti + i ,
(wT (xi ) + b) yi ti + i ,

wT
i w ti ,
ti 0, i 0, i 0, i = 1, . . . , N .
126
In order to utilize the Mercers kernel, we rst present the following theorem.
Theorem 6.2. If the corresponding local covariance
i can be estimated by
the mapped training data, i.e. i ,
i can be written as
i =
k

1
(xi+j ) ,
2k + 1
(6.17)
j=k
i =
k

1
((xi+j ) i )((xi+j ) i )T ,
2k + 1
(6.18)
j=k
where we just consider 2k data points which are the closest to the i-th data,
then the optimal w lies in the span of the mapped training data.
Proof. Suppose w = wp + wo , where wp is the projection of w in the span
of the mapped training data, wo is the orthogonal component to the span.
Since wT
rmo (xi ) = 0, i = 1, . . . , N , we can easily know that:
wT (xi ) = wT
p (xi ) ,
T
wT
i w = wp i wp .
Therefore, we can omit wo since it disappears in the optimization. We then

set it to 0 and obtain w = wp , i.e. the optimal w lies in the span of the
mapped training data.
N
By using Theorem 6.2, we write w as
j=1
j (xj ) and substitute it into
Eq.(6.17). By rewriting Eq.(6.17) in the kernel form by a kernel function

K(z 1 , z 2 ) = (z 1 )T (z 2 ), we then obtain:
wT (xi ) =
N

j K(xi , xj ) = T K i ,
j=1
i w
= T LT
i Li ,
where = [1 , . . . , N ]T , K i = [K(x1 , xi ) . . . K(xN

, xi )]T , K ij = K(xi ,
xj ),
K ik,1 . . . K ik,N
..
..
..
Li = 1 (K [ik:i+k,N ] 12k+1 lT
,
.
i ), K [ik:i+k,N ] =
.
.
2k+1
(lT
i )t =
1
2k+1
k
K i+k,1 . . . K i+k,N
K(xi+j , xt ), and 12k+1 is a column vector with ones of di-
j=k
mension 2k + 1.
Consequently, the corresponding objective in Eq.(6.16) becomes:
6.6 Additional Interpretation on w T i w

min
,b,ti ,i ,i
N
N

1
ti + C
(i + i )
N i=1
i=1
127

,
s.t. yi (T K i + b) ti + i ,
(T K i + b) yi ti + i ,

T LT
i Li ti ,
ti 0, i 0, i 0, i = 1, . . . , N .
Hence we only need a kernel function in the optimization without knowing a
specic mapping function and it can be easily solved by the SOCP methods.
6.6 Additional Interpretation on wT iw

We now interpret in terms of sparse approximation [2, 3, 7, 5, 4, 9, 14] why
wT i w can be considered as the local complexity around the data point xi .
In [7], Girosi has demonstrated an equivalence between sparse approximation and Support Vector Machines. In the view of sparse approximation,
the regression can be regarded as the task of approximating data using linear superpositions of basis functions selected from a large, redundant set of
basis functions, called dictionary [12]. A common sense in choosing a good
approximating function is that one should not only approximate the given
data as accurately as possible, more importantly, one should use as few as
possible basis functions. Therefore, a sparsity concept is invoked, i.e. the approximating function should be sparse in using the basis functions. When it
is connected with Support Vector Regressions, the readers can regard that
a basis function is associated with each data point (note that the regression function can be represented as the linear combination form in the kernel
space). The fact that SVR contains the property of sparsity, i.e. only a small
fraction of data points (support vectors) makes contributions to the nal
approximating function, may therefore explain why it has achieved a great
success. The measure of sparsity of the approximating function f , which is
also regarded as the measure of complexity is formulated as follows:
N p

[f ] =
i ,
(6.19)
i=1
where, i =
1, if xi appears ;
0, otherwise .
(6.20)
It is well known that the 0 -norm of a vector counts the number of elements
dierent from zero. The complexity term can also be described as:
[f ] = wp0 .
(6.21)
128
However, due to involving in minimizing a combinatorial term as the above,

it is extremely dicult to perform the optimization in practice. Therefore,
instead, one often uses 1 -norm as its approximated version, i.e.
[f ] = wp1 .
(6.22)
When p is set to 1, it therefore leads to the standard 1 -norm SVR. When

N

one looks back on the LSVR model, minimizing (1/N )
wT i w presents
i=1
another approximated version to the sparsity, since it also tries to make w as

N

sparse as possible.1 Another advantage of using (1/N )
wT i w is that
i=1
it leads to an easy solving method as illustrated in Section 6.4.
6.7 Experiments
In this section, we report the experiments on both synthetic sinc datasets and
real world datasets. The SOCP problem associated with our LSVR model is
solved by a general software, Sedumi [18, 19]. The SVR algorithm is performed by LIBSVM [1].
6.7.1 Evaluations on Synthetic Sinc Data
Fifty examples (xi , yi ) are generated from a sinc function [16], where xi are
drawn uniformly from [3, 3], and yi = sin(xi )/(xi ) + i , with i drawn
from a Gaussian with zero mean and variance 2 . Two cases are evaluated.
One is with = 0. The standard deviation of the data in the other case
increases linearly from 0.5 at x = 3 to 1.5 at x = 3. It is clearly observed that
in the second case, the variance of noise is dierent in dierent regions. We use
the default parameters C = 100, the RBF kernel K(u, v) = exp(u v2 ).
Table 6.1 reports the average results over 100 random trails with dierent
values. Fig. 6.2 illustrates the dierence between the LSVR model and the
SVR algorithm when = 0.2. For the case I, = 0.0, the LSVR model can
adjust the tube automatically to t the data with a smaller Mean Square
Error (MSE), which can be seen in Fig. 6.2(c). However, containing a xed
tube, the SVR algorithm lacks the exibility (see Fig. 6.2(a)). This also yields
that the MSE increases as increases. As reported in Table 6.1, when 0.8,
there are no support vectors in SVR and MSE is the largest. In case II, the
LSVR model has smaller MSEs and smaller STDs for all s. Fig. 6.2(d) also
shows that the obtained approximating function in LSVR is smoother than
that in SVR.
1
Intuitively, when w is sparser, (1/N )
N

i=1
w T i w would be smaller.
6.7 Experiments
129
Table 6.1. Experimental results (MSESTD) of the LSVR model and the SVR
algorithm on the sinc data with dierent values

Case I: = 0.0
LSVR
0.0
SVR
Case II: Varying

LSVR
SVR
0 0.18250.1011 0.31010.1165
0.2 0.0004
0.0160 0.23380.0888 0.27610.1111
0.4 0.0016
0.0722 0.19170.0726 0.22170.0840
0.6 0.0044
0.1695 0.15400.0687 0.23840.0867
0.8 0.0082
0.1748 0.13330.0674 0.23330.1096
1.0 0.0125
0.1748 0.11150.0597 0.25520.1218
2.0 0.0452
0.1748 0.09590.0421 0.26160.1517
(a) SVR, =0.0
(c) LSVR, =0.0
(b) SVR with varying
(d) LSVR with varying
Fig. 6.2. Experimental results on synthetic sinc data with =0.2
130
6.7.2 Evaluations on Real Financial Data

We evaluate our model on the nancial time series data which are highly
volatile and non-stationary. The experimental data are three major indices:
(1) the Dow Jones Industrial Average (DJIA), (2) the NASDAQ, and (3) the
Standard & Poor 500 index (S&P500) in the period from January 2, 2004
to April 30, 2004. We choose this period of data because three indices data
contain dierent statistical properties as reported in Table 6.2. Especially,
one may note that the data in this period for three indices contain largely
dierent skewness. In this way, the diversity in the data may not bias the
comparison of the models.
Table 6.2. Summary statistics of normalized returns of DJIA,
NASDAQ and S&P500 in the experiments. These indices show
dierent statistical properties.
Moments
DJIA
Train
Test
NASDAQ
Train
Test
S&P500
Train
Test
Mean
0.0000 0.2850 0.0000 0.4819
0.0000 0.3858
S.D.
1.0000
0.9957
1.0000
1.1312
1.0000
Skew
0.0678
0.1684
0.0928
0.3256 0.1298 0.0102
Kurt
2.5437
2.7706
2.6600
1.8631
2.5308
1.1298
2.4124
Following the procedure in [15], we convert the daily closing prices (dt )
of these indices to continuously compounded returns (rt = log(dt+1 /dt )) and
set the ratio of the number of the training return series to the number of
test return series to 5 : 1. We perform normalization on these return series
by Rt = (rt M ean(rt ))/SD(rt ), where the means and standard deviations
are computed for each individual index in the training period.
We compare the performance of the LSVR model against the SVR. The
t = f (xt ), where xt takes the previous four
predicted system is modelled as R
days normalized returns as indicators, i.e. xt = (Rt4 , Rt3 , Rt2 , Rt1 ).
Here this simple setting we employ is based on the suggestions in [15]: A
suitable selection for the sequent values is four. We then apply the modelled
function f to test the performance by one-step ahead prediction. The trade-o
parameter C and the parameter of the RBF kernel (K(u, v) = exp(u
v2 )), (C, ), are obtained by a ve-fold cross-validation conducting the
SVR on the following paired points: [25 , 24 , . . . , 210 ] [25 , 24 , . . . , 210 ].
We obtain the corresponding parameters as (24 , 23 ) for DJIA, (23 , 21 ) for
NASDAQ, and (20 , 22 ) for S&P500.
As suggested in [15], there is a relationship in the sequential ve days
values. We select k = 2, i.e. ve days values, to model the local volatility.
Since when 2.0, there are no support vectors in SVR, we just set the
References
131
values from 0.0, 0.2, . . . , 1.0, to 2.0. The corresponding results are reported
in Table 6.3. As observed, the LSVR model demonstrates a consistent superiority to the SVR algorithm, even though the paired parameters (C, ) are
not tuned for our LSVR model. Furthermore, a paired t-test [13] performed
on the best results of both models in Table 6.3, shows that the LSVR model
outperforms SVR with = 10% signicance level for a one-tailed test.
Table 6.3. Experimental results of the LSVR model and the SVR algorithm on the nancial data with dierent values

DJIA
NASDAQ
LSVR SVR
LSVR SVR
S&P500
LSVR
SVR
0.0 0.9204 1.3241 1.2897 1.3050 1.2372 1.2833

0.2 0.9835 1.1274 1.2896 1.3246 1.2399 1.2831
0.4 0.9341 0.9156 1.2898 1.3314 1.2442 1.2952
0.6 0.9096 0.9387 1.2901 1.3404 1.2540 1.2887
0.8 0.9273 0.9450 1.2904 1.3891 1.2788 1.2798
1.0 0.9434 0.9713 1.2908 1.4105 1.3044 1.2664
2.0 0.9666 1.0337 1.2928 1.3619 1.2643 1.3220
6.8 Summary
In this chapter, we propose a Local Support Vector Regression model. Different from the standard Support Vector Regression model, our novel model
oers a systematic and automatic scheme to locally and exibly adapt the
margin. Therefore, it can tolerate the noise adaptively. We demonstrate that
the promising model can not only capture the local information of the data
in approximating functions, but also can branch out similar models to the
standard SVR. The experiments conducted on sinc datasets and three indices
data from stock markets show that our model outperforms the standard SVR.
One future work of this model is to investigate ecient methods to directly
solve the original optimization of LSVR instead of solving a relaxed form. In
addition, both theoretical and empirical comparisons between the true solution and the approximated relaxed solution quantitatively are also valuable
research topics in the future.
References
1. Chang CC, Lin CJ (2001) LIBSVM: A Library for Support Vector Machines
2. Chen S (1995) Basis Pursuit. PhD thesis, Department of Statistics, Standford
University
132
References
3. Chen S, Donoho D, Saunders M (1995) Atomic decomposition by basis pursuit.

Technique Report 479, Department of Statistics, Standford University
4. Coifman RR, Wickerhauser MV (1992) Entropy-based algorithms for best-basis
selection. IEEE Transactions on Information Theory 38(2):713718
5. Daubechies I (1992) Ten lectures on wavelets. In CBMS-NSF Regional Conferences Series in Applied Mathematics. Philadelphia, PA: SIAM
6. Drucker H, Burges C, Kaufman L, Smola A, Vapnik VN (1997) Support Vector
Regression Machines. In Mozer Michael C, Jordan Michael I, Petsche Thomas,
editors, Advances in Neural Information Processing Systems. Cambridge, MA:
The MIT Press 9:155161
7. Girosi F (1998) An equivalence between sparse approximation and support
vector machines. Neural Computation 10(6):14551480
8. Gunn S (1998) Support vector machines for classication and regression. Technical Report NC2-TR-1998-030, Faculty of Engineering and Applied Science,
Department of Electronics and Computer Science, University of Southampton
9. Harpur GF, Prager RW (1996) Development of low entropy coding in a recurrent network. Networks 7:277284
locally and globally. In the 21st International Conference on Machine Learning
(ICML-2004)
11. Lobo M, Vandenberghe L, Boyd S, Lebret H (1998) Applications of second
order cone programming. Linear Algebra and Its Applications 284:193228
12. Mallat S, Zhang Z (1993) Matching pursuit in a time-frequency dictionary.
IEEE Transactions on Signal Processing 41(12):33973415
13. Montgomery Douglas C, Runger George C (1999) Applied statistics and probability for engineers. New York, NY: John & Wileys, 2nd edition
14. Olshausen BA, Field DJ (1996) Emergence of simple-cell receptive eld properties by learning a sparse code for natural images. Nature 381:607609
15. Pompe Bernd (2002) Mutual information and relevant variables for predictions. In Soo Abdols, Cao Liangyue, editors, Modelling and forecasting nancial data: techniques of nonlinear dynamics. Boston, MA: Kluwer Academic
Publishers 6192
16. Sch
olkopf B, Bartlett P, Smola A, Williamson R (1999) Shrinking the Tube:
A New Support Vector Regression Algorithm. In Kearns MS, Solla SA, Cohn
DA, editors, Advances in Neural Information Processing Systems. Cambridge,
MA: The MIT Press 11:330336
17. Smola A, Sch
olkopf B (1998) A tutorial on support vector regression. Technical
Report NC2-TR-1998-030, NeuroCOLT2
18. Sturm JF (1999) Using sedumi 1.02, a matlab toolbox for optimization over
19. Sturm JF (2000) Central region method. In Frenk JBG, Roos C, Terlaky T,
Zhang S, editors, High Performance Optimization. Kluwer Academic Publishers
157194
Springer-Verlag 2nd edition
21. Yang H, King I, Chan L, Huang K (2004) Financial Time Series Prediction Using Non-xed and Asymmetrical Margin Setting with Momentum in Support
Vector Regression. In Rajapakse JC, Wang L, editors, Neural Information Processing: Research and Development, Studies in Fuzziness and Soft Computing.
New York, NY: Springer-Verlag 152:334350
7
Extension III: Variational Margin Settings
within Local Data in Support Vector
Regression
In Chapter 6, we propose a Local Support Vector Regression Model to include

the local information of data. In this chapter, we consider another extension
of the Support Vector Regression (SVR) which also includes the local information of data for a specic application, i. e. nancial engineering. Both these
models are motivated from the local viewpoint of data.
SVR is derived from the Support Vector Machine which is based on
the principle of Structural Risk Minimization (SRM). Due to its solid
theoretical ground, SVR has been applied successfully in time series prediction [9, 10]. Usually, when SVR is applied in time series forecasting, it uses
the -insensitive loss function to measure the empirical risk. This loss function contains an margin. It not only measures the training error (empirical
risk), but also controls the sparsity of the solution (the number of support
vectors). When the width of -margin increases, it may tend to reduce the
number of support vectors. Extremely, a too wide margin may result in a
constant regression function. When the width of -margin decreases, it may
increase the number of support vectors. Ultimately, all the data points are
used for support vectors [19]. In this case, it may include the data noise in
seeking the regression function. Hence, setting the width of -margin is very
important. It aects the complexity and the generalization of the regression
function indirectly.
Normally, the setting of is xed, which is a kind of global setting.
However, in some applications, e. g. nancial engineering, the global setting
will not be an optimal choice. Since nancial data are usually volatile and
noisy, we extend the previous global margin setting to a variation one which
includes the local information of data.
In the following, we will rst describe the SVR model briey in Section 7.1.
We then indicate the problem of margin settings in Section 7.2. To solve the
problem of margin settings, we propose a general -insensitive loss function
for SVR in Section 7.3. We further aim at a specic application, i. e. nancial
engineering by introducing momentum and including GARCH model for the
134
7 Extension III: Variational Margin Settings within Local Data
variational margin settings in Section 7.4. After the detailed experimental

setup and experimental results in Section 7.5, we conclude the chapter with
discussions in Section 7.6.
7.1 Support Vector Regression

The aim of SVR is to nd a function f with parameters w and b by minimizing
the regression error as follows:

1
l(f (xi ), yi ) ,
w, w + C
2
i=1
N
Rreg (f ) =
(7.1)
where , denotes the inner product. This Euclidean norm w, w measures
the atness of the function f . Minimizing w, w will make the regression
function as at as possible [16].
The function f is then dened as
f (x, w, b) = w, (x) + b ,
(7.2)
where (x) : x , maps x X(Rd ) into a high (possible innite) dimensional space , and b R.
There are several loss functions which could be used to measure the regression error, e.g. squared loss function, Hubers loss function, -insensitive
loss function, etc. In SVR, the -insensitive loss function is used to measure
the loss [19] (illustrated in Fig. 7.1):

0,
if |y f (x)| < ;
l (y, f (x)) =
(7.3)
|y f (x)| , otherwise .
The advantage of this loss function is that it could aect the seeking of
regression function implicitly.
Fig. 7.1. Linear regression in the feature space by loss function
7.1 Support Vector Regression
135
To solve the minimization of Eq.(7.1) with loss function of Eq.(7.3) is

equivalent to solving the following constrained minimization problem:
min
(w, b, () ) =

1
(i + i ) ,
w, w + C
2
i=1
N
(7.4)
subject to
yi (w, (xi ) + b) + i ,
(w, (xi ) + b) yi + i ,
()
i
(7.5)
0.
Here and below, for every i, it ranges from 1 to N and () is a shorthand

implying both the variables with and without asterisks. i and i measure
the up error and down error for the sample point (xi , yi ), respectively, see
Fig. 7.1.
A standard method to nd the optimal solution of the above minimization problem in Eq.(7.4), further nding the function f in Eq.(7.2), is to
construct the dual problem of this optimization problem (primal problem)
by the Lagrange Method and to translate the (primal) minimization problem to maximize its dual function. Therefore, the optimization becomes a
Quadratic Programming (QP) problem as follows [19]:
min
Q(() ) =
1
(i i )(j j )(xi ), (xj )
2 i=1 j=1
N
N

( yi )i +
i=1
N

( + yi )i ,
(7.6)
i=1
subject to
N

(i i ) = 0,
()
[0, C] .
(7.7)
i=1
After solving this QP problem, we obtain the objective function as:

f (x) =
N

(i i )(xi ), (x) + b ,
i=1
where , are the Lagrange multipliers used to pull and push f towards
to the observation y. Those sample points (xi , yi ) with nonzero i or i are
called support vectors.
By using the trick of kernel function, one could dene the kernel function as the inner product of mapping function, i. e. K(x, z) = (x), (z).
Therefore, one only needs to specify a kernel function without considering the
136
mapping function or the feature space explicitly. The property of the kernel
function is that it should satisfy the Mercers Theorem [6, 14].
Four kernel functions are common used:
Linear function: K(xk , xl ) = xk , xl ;
Polynomial function with parameter d, K(xk , xl ) = (xk , xl + 1)d ;
Radial Basis Function (RBF) with parameter :
K(xk , xl ) = exp(xk xl 2 ) ,
(7.8)
Hyperbolic tangent: K(xk , xl ) = tanh(2xk , xl + 1) .
7.2 Problem in Margin Settings

Since the width of -margin holds the ability to aect the complexity and the
generalization of the regression function indirectly, it is very important to
seek an optimal value for dierent applications. Commonly, the is dicult
to control [10], as one does not know beforehand which one is able to t the
curve better.
Usually, there are several methods to deal with it. Firstly, most practitioners set the value of as a non-negative constant value just for convenience.
For example, in [18], they simply set the margin width to 0. This amounts to
the least modulus loss function. In other instances, the margin width has been
set to a very small value [5, 9, 20]. The second method is the cross-validation
technique, e. g. [4, 10]. It is usually too expensive in terms of computation. A
more ecient approach is to use another variant called -SVR [12, 13, 14, 15],
which determines by using another parameter . It is stated that may
be easier to specify than . Another approach by Smola, et al. [17] is to nd
the optimal choice of based on maximizing the statistical eciency of a
location parameter estimator. They showed that the asymptotically optimal
should be scaled linearly with the input noise of training the data, and this
was veried experimentally. Recently, a regularization path was proposed for
SVR to seek optimal parameters in [7, 21].
In nancial time series, however, the data are noisy and high volatile. The
xed margin setting is not suitable for this special application. We therefore
extend the xed margin setting to variational ones.
7.3 General
-insensitive Loss Function
First, we note that the margin in -insensitive loss function contains two
characteristics: xed and symmetrical. Based on these two characteristics, we
have proposed a general -insensitive loss function and classied the margin
into four cases in [22]: Fixed and Symmetrical Margin (FASM), Fixed and
7.3 General -insensitive Loss Function
137
Asymmetrical Margin (FAAM), Non-xed and Symmetrical Margin (NASM)

and Non-xed and Asymmetrical Margin (NAAM). Table 7.1 gives a simple
description of these four categories. FASM is equivalent to the margin in
-insensitive loss function, see Fig. 7.2(a). FAAM is divided into up margin
and down margin, each margin is xed but they are not equal (Fig. 7.2(b)).
While NASM is with equal up margin and down margin, but they are varied
with data (Fig. 7.2(c)). NAAM combines two characteristics of the margin
(Fig. 7.2(d)).
Table 7.1. Margin categories
Symmetrical
Asymmetrical
Fixed
FASM
FAAM
Non-xed
NASM
NAAM
Fig. 7.2. Four categories in general -insensitive loss function of SVR
In the following, we will derive the SV formula based on the general

-insensitive loss function. The general -insensitive loss function splits the
margin in the original -insensitive loss function into two parts: up margin
and down margin,
if d(xi ) < yi f (xi ) < u(xi );

0,
(7.9)
l (f (xi ) yi ) = yi f (xi ) u(xi ), if yi f (xi ) u(xi );
f (xi ) yi d(xi ), if f (xi ) yi d(xi ),
138
where d(xi ), u(xi ) 0, are two functions determining the down-margin and
up margin at point xi respectively. When d(x) and u(x) are both constant
functions and d(x) = u(x), Eq.(7.9) amounts to the -insensitive loss function
in Eq.(7.3) and we label it as FASM (Fixed and Symmetrical Margin). When
d(x) and u(x) are both constant functions but d(x) = u(x), this case is
labeled as FAAM (Fixed and Asymmetrical Margin). In the case of NASM
(Non-xed and Symmetrical Margin), d(x) = u(x) but are varied with the
data. The last case is with a non-xed and asymmetrical margin (NAAM)
where d(x) and u(x) are varied with the data and d(x) = u(x).
In the same way, we use the standard method to nd the solution of
Eq.(7.1) with the cost function of Eq.(7.9) as [19] and obtain:

N

1
min
(i + i ) ,
(7.10)
w, w + C
()
2
w,b,
i=1
subject to
yi w, (xi ) b u(xi ) + i ,
w, (xi ) + b yi d(xi ) + i ,
()
0.
Using the standard primal-dual method as above, we also obtain a QP problem as follows:
1
(i i )(j j )(xi ), (xj )
2 i=1 j=1
N
min (() ) =
N

i=1
(u(xi ) yi )i +
N

(d(xi ) + yi )i ,
(7.11)
i=1
subject to
N

(i i ) = 0, i , i [0, C] .
i=1
This QP problem is very similar to the original QP problem in Eq.(7.6),

therefore, we just need to modify the SMO algorithm a little bit to implement
this QP problem. Practically, we add a new data structure to store both
margins: up margin, u(x), and down-margin, d(x). This will not impact the
time complexity of the SVR algorithm; we just need more space linear to
the size of data points to store the corresponding margins. We modify the
LIBSVM from [5] to implement the SVR algorithm.
After solving this QP problem, we then obtain the regression function:
f (x) =
N

i=1
(i i )(xi ), (x) + b ,
(7.12)
7.4 Non-xed Margin Cases
139
where , are corresponding Lagrange multipliers also used to pull and

push f towards to the observation y.
The computation of b is exploited by the Karush-Kuhn-Tucker (KKT)
conditions. Here, they are:
i (u(xi ) + i yi + w, (xi ) + b) = 0 ,
i (d(xi ) + i + yi + w, (xi ) b) = 0 ,
and
(C i )i = 0 ,
(C i )i = 0 .
Therefore, b can be computed as follows:

yi w, (xi ) u(xi ), for i (0, C) ;
b=
yi w, (xi ) + d(xi ), for i (0, C) .
()
When no i
(0, C), methods e. g. [5] are used.
7.4 Non-xed Margin Cases

7.4.1 Momentum
In [23], we have focused on the case of NAAM. More specially, we have added
a momentum term in the margin setting. The margin is a linear combination
of the standard deviation and the momentum. The up margin and downmargin are set in the following forms:
u(xi ) = 1 (xi ) + (xi ),
d(xi ) = 2 (xi ) (xi ),
i = 1, . . . , N,
i = 1, . . . , N,
(7.13)
where (xi ) is the standard deviation of input xi , (xi ) is the momentum at

point xi , 1 , 2 are both positive constants and is a non-negative constant.
Therefore, the width of margin at point xi is:
W (xi ) = (1 + 2 ) (xi ) .
It is determined by (xi ) and the sum of 1 and 2 . Here we called 1 , 2
as the coecients of the margin width. We also called as the coecient of
momentum and we know that the margin setting of Eq.(7.13) includes the
case of NASM (when = 0).
From [22], when = 0 and (x) > 0, the up margin is larger than the
down-margin and we can under-predict the stock price. While = 0 and
140
(x) < 0, the up margin is smaller than the down-margin and we can overpredict the stock price. A simple illustration is shown in Fig. 7.3. Based on
these observations, in our prediction we assume that we are risk aversion, or
downside risk aversion. When the stock price reveals an uptrend, we know
that it will not be always up, so we tend to under-predict the stock prices
in this case. On the contrary, when the stock price goes down, we tend to
over-predict it. We add this information in the margin setting by controlling
the momentum term.
Fig. 7.3. Margin settings: dashed lines are the bounds of margins; dasheddotted lines are actual data series; solid-bold lines are the new objective
function, f new , by new margin settings. The upper shadow area is the case
of new objective function under-predicted to the actual function; the lower
shadow parts are the case of over-predicted
Actually, there are many ways to calculate the momentum. For example,
the simplest way is to set it as a constant. In this chapter, we will concentrate
on using the Exponential Moving Average (EMA). The reason of using EMA
is that it is time-varying and can reect the uptrend and down-tendency of
the nancial data. A little deciency is that there exists the lag problem. An
n-days EMA sequence begins from the rst day, i. e. EM A1 = y1 and the
following is calculated by:
EM Ai = EM Ai1 (1 r) + yi r ,
where r = 2/(1 + n), and yi is the information about day i, e.g. the closing
price in day i, the volume in day i, etc. Here, the current days momentum
is set as the dierence between the current days EMA and the EMA in the
previous k day, i. e.
(xi ) = EM Ai EM Aik .
7.4.2 GARCH
In the above methods, the datasets we used in the experiments are the price of
the share [22, 23]. We use the standard deviation of input xt , which can reect
7.5 Experiments
141
the volatility of the nancial time series over time, to determine the width of
margin at time t in our prediction. Actually, the Generalized AutoRegressive
Conditionally Heteroscedastic (GARCH) model [3] is a more common used
model to reect the volatility of the nancial time series.
The standard GARCH(p, q) model with Gaussian shocks takes the following form:
yt = c0 + xT
t |t1 = N (0, t2 ) ,
t b + t ,
where
t2 = 0 +
p

i=1
2
i ti
+
q

j 2tj .
j=1
This GARCH toolbox is applied to the return series. So we use the continuous compounded return as the data series and use the t calculated by
GARCH(1,1) as the width of margin at time t.
7.5 Experiments
In this section, we will perform the experiments by using the momentum and
GARCH models to set the margins. Before illustrating the experiments, we
dene the accuracy and risk measurement rst.
7.5.1 Accuracy Metrics and Risk Measurement
In order to measure the prediction performance of our model, we dene the
Mean Absolute Error (MAE).
Let at and pt be the actual values respectively and predicted values at
day t, let m be the number of testing data.
Denition 7.1. Mean Absolute Error (MAE) measures the discrepancy
between the actual and predicted values; the smaller the value of MAE, the
closer are the predicted values to the actual values. MAE is calculated by:
MAE =
m
1
|at pt | .
m t=1
(7.14)
We also consider the risk of using this model in the prediction. Actually,
risk is a term frequently encountered in strategic management and nancial
literature. However, risk has a variety of dierent meanings and rarely is
the meaning used in a particular project claried in [2]. In nancial literature, Markowitz rst formulated the portfolio selection into a mathematical
model [8]. In his model, the return of a portfolio is measured by the expected value of the random portfolio return and the associated risk is quantied by the variance of the portfolio return. However, the use of variance
to measure risk makes no distinction between gains and losses. Markowitz
142
also proposed to use semi-variance to measure the risk of loss. That is the
sum of the squares of negative deviations from the mean divided by the total
number of observations:
1
[min(rt , 0)]2 .
m t=1
m
The great advantage of the use of semi-variance over variance is that

it does not include positive gains, so what is considered as risk takes into
account only negative deviations. However, minimizing downside does not
mean minimizing only negative deviations. For example, if the distribution,
like the normal curve, is symmetric, minimizing variance and semi-variance
will lead to the same problem. The only case that justies the use of semivariance is when the presence of skewness is observed [1]. A generalization of
semi-variance is given in [1]:
1
[min(rt , 0)]k ,
m t=1
m
downside risk
(7.15)
where k is any power that one chooses; when k=1, it should be considered
the absolute value of the term in the brackets and is a chosen benchmark
(not necessarily the mean).
Based on Eq.(7.15), we choose k=1 and dene the following risk measurements.
Denition 7.2. Upside Mean Absolute Error (UMAE) measures upside risk; the smaller the value of UMAE, the smaller the upside risk. UMAE
is dened as:
m
1
UMAE =
(at pt ) .
(7.16)
m t=1
at pt
Denition 7.3. Downside Mean Absolute Error (DMAE) measures

the downside risk; the smaller the value of DMAE, the smaller the downside
risk. DMAE is dened as:
DMAE =
m
1
(pt at ) .
m t=1
(7.17)
at <pt
7.5.2 Momentum
We compare the modied SVR algorithm by adapting margins using momentum with the AutoRegression (AR) model and the Radial Basis Function
(RBF) method. The results are presented as follows one by one for three
algorithms.
7.5 Experiments
143
7.5.2.1 SVR Algorithm

Two datasets are used in this experiment:
HSI: daily closing prices of Hong Kongs Hang Seng Index (HSI) from
January 2nd, 1998 to December 29, 2000.
DJIA: daily closing prices of Dow Jones Industrial Average (DJIA) from
January 2nd, 1998 to December 29, 2000.
The ratio of the number of training data and the number of testing data
is set to 5:1. Therefore, the corresponding initial training time periods are
obtained and listed as in Table 7.2.
Table 7.2. Indices, time periods and parameters for momentum experiments
Indices
Initial training time periods
HSI
02/01/1998 04/07/2000
16000
227
DJIA
02/01/1998 29/06/2000
8000
222
Furthermore, we model the system as pt = f (xt ), where f is learned by

the SVR algorithm from the training data, xt = (at4 , at3 , at2 , at1 ), at
is the daily closing index in day t.
Before generating the model, we do a cross-validation on the initial training data to determine the parameters that are needed in SVR. They are C,
the cost of error and , the parameter of kernel function. The corresponding
parameters are also listed in Table 7.2. With these parameters we begin to
build the model by SVR from the initial training data. After obtaining the
predictive value, we shift the input window to the next time-step and train
the model again to obtain the next days price. This one-step ahead prediction
is done as the window shifted for the remaining data.
Non-xed Cases: The margins setting is followed as Eq.(7.13). In the
case of NASM, we set 1 = 2 = 1/2 and = 0, thus the overall margin
width at day t is equal to the standard deviation of input xt , (xt ).
In the case of NAAM, we also x 1 = 2 = 1/2, hence we have a fair comparison of NASM case. In addition, we have to determine three parameters,
i.e. n, the length of EMA; k, the lag of EMA; , the coecient of momentum.
We have performed the following experiments to test their eects:
(a) At rst, we set k = 1, = 1 and use 10, 30, 50, 100 as the length of
EMA respectively. From the result of Table 7.3 we can see that the DMAE
values in all cases of NAAM are smaller than that in NASM case, thus we have
a smaller downside risk in NAAM case; this exactly meets our assumption.
We also see that the MAE gradually decreases with the increase of the length
of EMA, and that when the length equals 100, the MAE and the DMAE are
144
the smallest in all cases of NAAM for dataset HSI. For dataset DJIA, when
the length equals 30, the MAE and the DMAE are also the smallest in all
cases of NAAM.
Table 7.3. Eect of the length of EMA on HSI with parameters
(k, )=(1,1)
Type
DJIA
MAE
UMAE DMAE
MAE UMAE DMAE
216.78
104.58
112.20
85.33
40.29
45.04
10
222.43
115.64
106.79
85.68
43.13
42.55
30
218.18
114.04
104.14
84.12
41.82
42.30
50
217.93
113.38
104.55
84.57
42.12
42.45
100 216.50
113.04
103.46
84.80
42.41
42.39
NASM
NAAM
HSI
In the following, we will use the best length of EMA from the above
experiments for the corresponding datasets, i. e. n = 100 for data set HSI
and n = 30 for dataset DJIA.
(b) When testing the eect of lag k, we let = 1 and set k to 1, 2, 4, 8
respectively for both datasets. The results are listed in Table 7.4. They show
that the MAE increases with increasing of the lag of EMA. These indicate
that the results when the lag of EMA equals 1 are superior to the other cases.
Table 7.4. Eect of the distance of EMA on HSI and DJIA
HSI with (n, k) = (100, 1)
DJIA with (n, k) = (30, 1)
MAE
UMAE
DMAE
MAE
UMAE
DMAE
216.50
113.04
103.46
84.12
41.82
42.30
219.02
125.30
93.72
85.42
43.91
41.51
228.25
149.36
78.88
90.99
49.16
41.83
260.73
200.74
59.99
103.77
58.03
45.74
(c) Here, we set k = 1 and = 1, 1/2, 1/4, 1/8 respectively for both
datasets to see the eect of the . From Table 7.5, we see that the DMAE
increases gradually with decreasing of the coecient of EMA and that the
MAE is smaller than the value in the NASM case. The change of the MAE
for dataset HSI in (24 columns of) Table 7.5 is uctuating and the MAE
7.5 Experiments
145
in (57 columns of) Table 7.5 increases gradually with the decrease of the
coecient of EMA.
Table 7.5. Eect of the coecient of momentum on HSI and DJIA
HSI with (n, k) = (100, 1)
DJIA with (n, k) = (30, 1)
MAE
UMAE
DMAE
MAE
UMAE
DMAE
216.50
113.04
103.46
84.12
41.82
42.30
1/2
216.55
108.97
107.58
84.88
41.32
43.56
1/4
216.19
106.36
109.83
85.02
41.14
43.88
1/8
216.41
105.32
111.08
85.22
40.86
44.36
We also plot the daily closing prices of HSI with 100 days EMA and
the prices of DJIA with 30 days EMA in Fig. 7.4 and Fig. 7.5 respectively,
and list the Average Standard Deviations (ASD) of input x of the training
datasets HSI and DJIA, respectively in Table 7.6, the Average of Absolute
Momentums (AAM) of input x for the best length of both training datasets
respectively in Table 7.6. We can observe that the ASD of HSI is higher than
that of DJIA and that the ratio of AAM to ASD is smaller for HSI than that
for DJIA.
Table 7.6. ASD and AAM
AAM
Dataset
ASD
HSI
DJIA
Ratio
182.28
100
20.80
0.114
79.95
30
15.64
0.196
Now, we will make a summary for the above experiments. At rst, we can
know the eects of n, k and from the above experiments results. Following
these results, we can say that a suitable setting for k and will both be
1, which can be applied when a new dataset comes. The only parameter
needed to determine is the length of EMA, n, this may refer to the ASD of
the training dataset. When the ASD is larger, we may use a longer length
of EMA. On the contrary, when the ASD is smaller, we may use a shorter
length of EMA.
Fixed Cases: After considering the non-xed margin cases, we also test
the predictive results of xed margins. Actually, for dataset HSI, we let
146
Fig. 7.4. HSI with 100 days EMA
Fig. 7.5. DJIA with 30 days EMA
the width of margin equal to 200 (approximate to the ASD of HSI), i. e.

u(x) + d(x) = 200. The up-margin u(x) ranges from 0 to 200, each increment is one-tenth of 200, i. e. 20. The results are listed in (15 columns of)
Table 7.7. Similarly, for dataset DJIA, we let the width of margin equal to
90 (approximate to ASD of DJIA), i. e. u(x) + d(x) = 90. The up-margin
7.5 Experiments
147
u(x) ranges from 0 to 90, each increment is also one-tenth of 90, i. e. 9. The
results are listed in (610 columns of) Table 7.7. We can see that for both
datasets, as the up-margin increases, the DMAE tends to decrease.
Table 7.7. Results of FASM and FAAM for HSI and DJIA
HSI [u(x)+d(x)]
DJIA [u(x)+d(x)]
u(x) d(x) MAE UMAE DMAE
u(x) d(x) MAE UMAE DMAE
200 236.04
62.24
173.80
90
91.63
20.45
71.18
20
180 230.85
69.65
161.20
81
89.14
23.70
65.44
40
160 226.29
77.37
148.92
18
72
87.35
27.31
60.04
60
140 222.24
85.34
136.90
27
63
86.09
31.18
54.91
80
120 219.35
93.90
125.45
36
54
85.30
35.28
50.02
100
100 217.83 103.14
114.69
45
45
85.45
39.86
45.59
120
80
217.35 112.90
104.45
54
36
86.33
44.80
41.53
140
60
217.88 123.16
94.72
63
27
87.40
49.83
37.57
160
40
219.49 133.97
85.52
72
18
88.64
54.95
33.69
180
20
221.66 145.05
76.61
81
90.80
60.53
30.27
200
224.83 156.64
68.19
90
93.75
66.51
27.24
Comparing the results in Table 7.3 with the results in Table 7.7 (the
experimental results are plotted in Fig. 7.6(b) and Fig. 7.7(b) respectively),
we can see that NASM and NAAM are both superior to FASM and FAAM
in both datasets.
In the following, we will perform other models, such as AR models and
RBF network, on the above two datasets. The best results of all the models
are illustrated in Fig. 7.6(a) for HSI and Fig. 7.7(a) respectively.
7.5.2.2 AR Models
For AR models, we use the AR model with order 4 to predict the prices of
HSI and DJIA, hence we can compare the AR model with NASM, NAAM in
SVR with the same order. The results are listed in the Table 7.8. From these
results, we can see that NASM and NAAM are superior to AR model with
the same order.
148
Fig. 7.6. Experimental results of HSI
Fig. 7.7. Experimental results of DJIA

Table 7.8. Results on AR(4)
Dataset
MAE
UMAE
DMAE
HSI
217.75
105.96
111.79
DJIA
88.74
46.36
42.38
7.5.2.3 RBF Network

For the RBF network, we use the RBF network which was implemented in
NETLAB [11] and perform the one-step-ahead prediction to predict the prices
of HSI and DJIA. Concretely, we let other parameters as default and set the
number of hidden units to 3, 5, 7, 9 to learn f by training the RBF network
on the training samples, and obtain the results in Table 7.9 for both datasets.
Comparing the results in Table 7.3 with the results in Table 7.9, we can see
that NASM and NAAM are also better than RBF network.
7.5 Experiments
149
Table 7.9. Eect of number of hidden units on HSI and DJIA

HSI
Hidden No.
DJIA
MAE UMAE DMAE MAE UMAE DMAE

3
386.65 165.08 221.57
88.31 44.60
43.71
277.83 128.92 148.91
98.44 48.46
49.98
219.32 104.15 115.17
90.53 46.22
44.31
221.81 109.46 112.35
87.23 44.09
43.14
7.5.3 GARCH
In this experiment, the experimental data are 3 years daily closing indices
(20002002) from stock markets in dierent countries:
Nikkei225: Nikkei225 Stock Average from Japan, the daily closing prices
are plotted in Fig. 7.11(a);
DJIA00-02: Dow Jones Industrial Average (DJIA) from USA, the daily
closing prices are plotted in Fig. 7.13(a);
FTSE100: FTSE100 index from UK, the daily closing prices are plotted
in Fig. 7.15(a).
In the data processing step, the daily closing prices of these indices are
converted to continuously compounded returns and the ratio of the number
of training data to the number of testing data is set to 5:1. Therefore, we
obtain and list the corresponding training and testing periods in Table 7.10.
Table 7.10. GARCH experimental data description
Indices
Training period
Testing period
Nikkei225
4 Jan., 2000 2 Jul., 2002
4 Jul., 2002 30 Dec., 2002
DJIA00-02
3 Jan., 2000 3 Jul., 2002
5 Jul., 2002 31 Dec., 2002
FTSE100
4 Jan., -2000 3 Jul., 2002
4 Jul., 2002 31 Dec., 2002
7.5.3.1 GARCH(1, 1)
We apply the Matlab toolbox to calculate the GARCH model. In the Matlab
toolbox, Before running the SVR algorithm, we run the GARCH(1,1) model
to determine the width of margin in SVR. For Nikkei225, we obtain the
parameter estimates and their standard errors in Table 7.11, i. e. the best ts
for Nikkei225 by (1,1) is:
150
yt = 0.49468 + t ,
2
+ 0.0772182t1 .
t2 = 0.00073917 + 0.8682t1
Table 7.11. GARCH parameter for Nikkei225

Standard
Parameter
Value
c0
0.49468
0.0045008
109.9083
0.00073917
0.00034866
2.1200
GARCH(1)
0.8682
0.048144
18.0334
ARCH(1)
0.077218
0.027279
2.8306
error
statistic
We also show that the log-likelihood contours of GARCH(1,1) model

t to the returns of dataset, Nikkei225 Fig. 7.8(a) The log-likelihood contours are plotted in a GARCH coecient-ARCH coecient (G1 A1 ) plane,
holding the parameters c0 and 0 xed at their maximum likelihood estimates 0.49468 and 0.00073917, respectively. The contours conrm the results
in Table 7.11. The maximum log-likelihood value occurs at the coordinates
G1 = GARCH(1) = 0.8682 and A1 = ARCH(1) = 0.077218. This gure also
reveals a highly negative correlation between the estimates of the G1 and
A1 parameters of the GARCH(1,1) model. It implies that a small change in
the estimate of the G1 parameter is nearly compensated for a corresponding
change of opposite sign in the A1 parameter. The innovations, standard deviations (t ) and returns of Nikkei225 are shown in Fig. 7.8(b).
Fig. 7.8. GARCH(1,1) of Nikkei225. The color-coded bar at the right of (a) indicates the height of the log-likelihood surface of the GARCH(1,1) plane
7.5 Experiments
151
For dataset DJIA00-02, GARCH(1,1) parameter estimates are listed in

Table 7.12, therefore, the best ts for DJIA00-02 by GARCH(1,1) is
yt = 0.60363 + t ,
2
+ 0.0922952t1 .
t2 = 0.00056832 + 0.85971t1
Table 7.12. GARCH parameter for DJIA00-02

Standard
Parameter
Value
c0
0.60363
0.0041185
146.5631
0.00056832
0.00023491
2.4193
GARCH(1)
0.85971
0.031773
27.0580
ARCH(1)
0.092295
0.020352
4.5350
error
statistic
The corresponding log-likelihood contours of DJIA00-02 are plotted in

Fig. 7.9(a), the maximum log-likelihood value occurs at the coordinates
G1 = GARCH(1) = 0.85971 and A1 = ARCH(1) = 0.09229. The corresponding innovations, standard deviation and returns of DJIA00-02 are
shown in Fig. 7.9(b).
Fig. 7.9. GARCH(1,1) of FTSE100. The color-coded bar at the right of (a) indicates the height of the log-likelihood surface of the GARCH(1,1) plane
For dataset FTSE100, GARCH(1,1) parameter estimates are listed in

Table 7.13 therefore, the best ts for FTSE100 by GARCH(1,1) is
152
yt = 0.50444 + t ,
2
+ 0.126932t1 .
t2 = 0.0011599 + 0.82253t1
Table 7.13. GARCH parameter for FTSE100

Standard
Parameter
Value
c0
0.50444
0.0053313
error
T
statistic
94.6180
0.0011599
0.00049206
2.3573
GARCH(1)
0.82253
0.04906
16.7658
ARCH(1)
0.12693
0.034698
3.6582
The corresponding log-likelihood contours of FTSE100 are plotted in

Fig. 7.10(a). The maximum log-likelihood value occurs at the coordinates
G1 = GARCH(1) = 0.82253 and A1 = ARCH(1) = 0.12693. The corresponding innovations, standard deviation and returns of FTSE100 are shown
in Fig. 7.10(b).
Fig. 7.10. GARCH(1,1) of DJIA00-02. The color-coded bar at the right of (a)
indicates the height of the log-likelihood surface of the GARCH(1,1) plane
7.5.3.2 SVR Algorithm

For SVR algorithm, the experimental procedure consists of three steps: at
rst, we normalize the return value by ti = (ri rlow )/(rhigh rlow ), where ri
is the actual return of the stock at day i, rlow and rhigh are the correspondingly minimum and maximum return in the training data, respectively. Then,
7.5 Experiments
153
we train the normalized training data once and then obtain the normalized
predicted return value pni = f (xi ), where xi = (ti4 , ti3 , ti2 , ti1 ). Finally,
we unnormalize pni , convert the result to price and obtain the corresponding
predicted price pi .
Before running the SVR algorithm, we have to choose two parameters: C,
the cost of error; , the parameter of kernel function. Here the parameters
we choose are the same respectively for dierent indices. They are listed in
Table 7.14.
Table 7.14. Parameters in GARCH experiments for NASM
Indices
Nikkei225
24
DJIA
24
FTSE100
24
Here, we just consider the case of NASM. The margin setting is as

Eq.(7.13). Concretely, we set the margin width to calculated by GARCH(1,1)
from return series y, therefore 1 = 2 = 1/2 and = 0. For xed
margin cases, we set the margin width as 0.1, i. e. u(x) + d(x) = 0.1,
and each increment is 0.02. The corresponding results are shown in the
Tables 7.157.17. We also plot the training and testing data results of NAAM
in Figs. 7.12(a) and 7.12(b) for index Nikkei225, in Figs. 7.14(a) and 7.14(b)
for index DJIA00-02, in Figs. 7.16(a) and 7.16(b) for index FTSE100, respectively. From these results, we can see that for FTSE100 index, NASM
outperforms in the prediction than in xed margin cases. For Nikkei225, when
u(x) = 0.06, d(x) = 0.04 and u(x) = 0.08, d(x) = 0.02, the predicted results
are better than NASM. For DJIA00-02, when u(x) = 0.06, d(x) = 0.04, the
predicted result is slightly better than NASM.
7.5.3.3 AR Models
We also use AR model with dierent orders (16) to predict the prices of the
above three indices. The experimental procedure is to apply the AR model on
training return series and to obtain the predicted return value from testing
data. Then we convert the predicted return values to price values. We obtain
the experimental results and show them in Table 7.18. After comparing the
results in Tables 7.15 and 7.17 with the results in 24 and 810 columns
of Table 7.18, we can see that for Nikkei225 and FTSE100 index, the NASM
method is better than AR model. For DJIA, we can see that NASM method
is slight worse than AR(1), but better than other order of AR model.
For index Nikkei225, the predictive error and risks comparison results
graphs are shown in Fig. 7.11(b), the corresponding bar values are from
154

Table 7.15. SVR results for Nikkei225
Type
u(x)
d(x)
MAE
UMAE
DMAE
NASM
124.37
55.97
68.40
0.10
141.60
30.70
110.90
0.02
0.08
131.25
39.02
92.23
0.04
0.06
125.63
49.66
75.97
0.06
0.04
123.11
61.81
61.30
0.08
0.02
124.00
75.63
48.37
0.10
129.19
91.56
37.63
FAAM
Table 7.16. SVR results for DJIA00-02

Type
u(x)
d(x)
MAE
UMAE
DMAE
NASM
129.56
62.74
66.83
0.10
139.82
41.56
98.26
0.02
0.08
134.33
49.16
85.17
0.04
0.06
130.49
57.56
72.93
0.06
0.04
128.51
66.87
61.64
0.08
0.02
129.65
77.72
51.94
0.10
133.76
90.02
43.74
FAAM
Table 7.17. SVR results for FTSE100

Type
u(x)
d(x)
MAE
UMAE
DMAE
NASM
69.61
33.42
36.19
0.10
73.46
25.93
47.53
0.02
0.08
71.98
28.52
43.46
0.04
0.06
70.83
31.27
39.56
0.06
0.04
70.10
34.22
35.88
0.08
0.02
69.86
37.42
32.45
0.10
70.26
40.92
29.34
FAAM
7.6 Discussions
155
Table 7.18. AR results
Order
Nikkei225
DJIA00-02
FTSE100
MAE UMAE DMAE MAE UMAE DMAE MAE UMAE DMAE

1
125.31 53.40
71.91
128.58 61.67
66.91
71.44 33.9
37.53
125.68 53.31
72.36
130.00 62.08
67.92
71.40 33.46
37.94
125.67 53.37
72.30
130.56 62.50
68.06
70.41 32.76
37.65
125.22 52.91
72.31
131.20 62.93
68.27
69.96 32.76
37.20
125.32 53.08
72.24
131.27 62.90
68.38
70.12 32.89
37.23
125.40 52.72
72.68
131.32 62.89
68.43
69.99 32.78
37.21
Table 7.15 and (24 columns of) Table 7.18. The predictive error and risks
of DJIA00-02 are shown in Fig. 7.13(b), where the corresponding bar values
are from Table 7.16 and (57 columns of) Table 7.18. The predictive error
and risks of FTSE100 are shown in Fig. 7.15(b), where the corresponding bar
values are from Table 7.17 and (810 columns of) Table 7.18.
Fig. 7.11. Nikkei225 data plot and experimental results graphs
7.6 Discussions
Having described the experiments and their results, we know that NASM is
superior to FASM and FAAM generally. One reason is that NASM catches
the stock market information and adds the information into the setting of the
156
Fig. 7.12. Experimental results graphs using GARCH method for Nikkei225
Fig. 7.13. DJIA00-02 data plot and experimental results graphs
Fig. 7.14. Experimental results graphs using GARCH method for DJIA00-02
7.6 Discussions
157
Fig. 7.15. FTSE100 data plot and experimental results graphs
Fig. 7.16. Experimental results graphs using GARCH method for FTSE100
margin. This provides helpful information for the prediction. Another reason
is that by using NASM, the margin width is determined by a meaningful
value. This value changes with the stock market. Obviously, this method is
more exible than xed margin cases and avoids risk of getting bad predictive
results partially when the margin values are determined by random selection
in the xed margin cases.
Furthermore, we know that NAAM may be better than NASM. For
example, by adding a momentum, we may not only improve the accuracy
of prediction, but also reduce the predictive downside risk.
Another notice is that by cautiously selecting parameters, SVR algorithm
has similar predictive performance to other models, from Figs. 7.6(a) and
7.7(a). However, for a novice, the SVR libraries are easy to run. Since every
local optimum is the global optimum, it guarantees the user to nd an optimal
solution easily and stably. This advantage is very useful for a novice to learn
a new model, or library, and strengthen his condence of learning new things
comparing with learning other non-linear model, e. g. RBF networks.
158
References
In general, our methods can be considered as a model selection, determining the parameter, . We do not consider the setting of other parameters,
such as C and . We just use the cross-validation technique to nd suitable
values for them. However, this procedure is time-consuming. We may add
some market information to set these parameters, e. g. [4]. In addition, the
margin width set by GARCH model is too wide; we may need to add more
useful terms to shrink it. This can be one of our future works. A valuable
experience is that the normalized procedure will be helpful for selecting suitable parameters easily and stably.
Finally, we turn to a key weakness of our model: the predictive model
does not lead to direct prot making in real life and we do not provide the
condence of these predictive models. However, we may nd some useful
information through using our model to predict the stock market prices; the
predictive results may provide some helpful suggestions.
References
1. Gustavo M, de Athayde (2001) Building a Mean-downside Risk Portfolio Frontier. In: Sortino F.A, Satchell S.E, editors, Managing Downside Risk in Financial Markets: Theory, Practice and Implementation. Oxford, Boston:Butterworth-Heinemann 194211
2. Baird IS, Howard T (1990) What Is Risk Anyway? Using and Measuring Risk
in Strategic Management. In Bettis Richard A and Thomas Howard, editors,
Risk, Strategy and Management. Greenwich, Conn: JAI Press 2151
3. Bollerslev T (1986) Generalized Autoregressive Conditional Heteroskedasticity.
Econometrics 31:307327
4. Cao LJ, Chua KS, Guan LK (2003) c-Ascending Support Vector Machines for
Financial Time Series Forecasting. In International Conference on Computational Intelligence for Financial Engineering (CIFEr2003) 329335
5. Chang CC, Lin CJ (2001) LIBSVM: A Library for Support Vector Machines
6. Cristianini N, Shawe-Taylor J (2000) An Introduction to Support Vector Machines(and Other Kernel-based Learning Methods). Cambridge, U.K.; New
York: Cambridge University Press
7. Hastie T, Rosset S, Tibshirani R, Zhu J (2004) The entire regularization path
for the support vector machine. Journal of Machine Learning Research 5:1391
1415
8. Markowitz H (1952) Portfolio Selection. Journal of Finance 7:7791
9. Mukherjee S, Osuna E, Girosi F (1997) Nonlinear Prediction of Chaotic Time
Series Using Support Vector Machines. In Principe J, Giles L, Morgan N,
Wilson E, editors, IEEE Workshop on Neural Networks for Signal Processing
VII. IEEE Press 511519
10. M
uller KR, Smola A, R
atsch G, Sch
olkopf B, Kohlmorgen J, Vapnik V (1997)
Predicting Time Series with Support Vector Machines. In Gerstner W, Germond A, Hasler M, and Nicoud JD, editors, ICANN. New York, NY: Springer
9991004
11. Nabney IT (2002) Netlab: Algorithms for Pattern Recognition. New York, NY:
Springer
References
159
12. Sch
olkopf B, Chen PH, Lin CJ (2003) A Tutorial on -Support Vector Machines. Technical Report, National Taiwan University
13. Sch
olkopf B, Bartlett P, Smola A, Williamson R (1998) Support Vector Regression with Automatic Accuracy Control. In Niklasson L, Boden M, and Ziemke
T, editors, Proceedings of ICANN98 Perspectives in Neural Computing. Berlin
111116
14. Sch
olkopf B, Bartlett P, Smola A, Williamson R (1999) Shrinking the Tube:
A New Support Vector Regression Algorithm. In Kearns MS, Solla SA, Cohn
DA, editors, Advances in Neural Information Processing Systems. Cambridge,
MA: The MIT Press 11: 330336
15. Sch
olkopf B, Smola AJ, Williamson R, Bartlett P (1998) New Support Vector Algorithms. Technical Report NC2-TR-1998-031, GMD and Australian
National University
16. Smola A, Sch
olkopf B (1998) A tutorial on support vector regression. Technical
Report NC2-TR-1998-030, NeuroCOLT2
17. Smola AJ, Murata N, Sch
olkopf B, M
uller KR (1998) Asymptotically Optimal
Choice of -Loss for Support Vector Machines. In Proc. of Seventeenth Intl.
Conf. on Articial Neural Networks
18. Trafalis TB, Ince H (2000) Support Vector Machine for Regression and Applications to Financial Forecasting. In Proceedings of the IEEE-INNS-ENNS
International Joint Conference on Neural Networks (IJCNN2000). IEEE 6: 348
353
Springer, 2nd edition
20. Vapnik VN, Golowich S, Smola AJ (1997) Support Vector Method for Function Approximation, Regression Estimation and Signal Processing. In Mozer
M, Jordan M, Petshe T, editors, Advances in Neural Information Processing
Systems. Cambridge, MA: The MIT Press 9: 281287
21. Wang G, Yeung DY, Lochovsky FH (2006) Two-dimensional solution path for
support vector regression. In The 23rd International Conference on Machine
Learning. Pittsburge, PA: 19931000
22. Yang H, Chan L, King I (2002) Support Vector Machine Regression for Volatile
Stock Market Prediction. In Yin Hujun, Allinson Nigel, Freeman Richard,
Keane John, and Hubbard Simon , editors, Intelligent Data Engineering and
Automated Learning IDEAL 2002. NewYork, NY: Springer 2412 of LNCS:
391396
23. Yang H, King I, Chan L (2002) Non-xed and Asymmetrical Margin Approach
to Stock Market Prediction Using Support Vector Regression. In International
Conference on Neural Information Processing ICONIP 2002, 1968
8
Conclusion and Future Work
In this chapter, a summary of this book is provided. We will review the whole
journey of this book, which starts from two schools of learning thoughts
in the literature of machine learning, and then motivate the resulting combined learning thought including Maxi-Min Margin Machine, Minimum Error
Minimax Probability Machine and their extensions. Following that, we then
present both future perspectives within the proposed models and beyond the
developed approaches.
8.1 Review of the Journey

Two paradigms exist in the literature of machine learning. One is the
school of global learning approaches; the other is the school of local learning
approaches. Global learning enjoys a long and distinguished history, which
usually focuses on describing phenomena by estimating a distribution from
data. Based on the estimated distribution, the global learning methods can
then perform inferences, conduct marginalizations, and make predictions.
Although containing many good features, e.g. a relatively simple optimization and the exibility in incorporating global information such as structure
information and invariance, etc., these learning approaches have to assume a
specic type of distribution a prior. However, in general, the assumption itself
may be invalid. On the other hand, local learning methods do not estimate
a distribution from data. Instead, they focus on extracting only the local
information which is directly related to the learning task, i.e. the classication in this book. Recent progress following this trend has demonstrated that
local learning approaches, e.g. Support Vector Machine (SVM), outperform
the global learning methods in many aspects. Despite of the success, local
learning actually discards plenty of important global information on data,
e.g. the structure information. Therefore, this restricts the performance of
this types of learning schemes. Motivated from the investigations of these
162
8 Conclusion and Future Work
two types of learning approaches, we therefore suggest to propose a hybrid

learning framework. Namely, we should learn from data globally and locally.
Following the hybrid learning thought, we thus develop a hybrid model
named Maxi-Min Margin Machine (M4 ), which successfully combines two
largely dierent but complementary paradigms. This new model is demonstrated to contain both appealing features in global learning and local learning. It can capture the global structure information from data, while it can
also provide a task-oriented scheme for the learning purpose and inherit the
superior performance from local learning. This model is theoretically important in the sense that M4 contains many important learning models as
special cases including Support Vector Machines, Minimax Probability Machine (MPM), and Fisher Discriminant Analysis; the proposed model is also
empirically promising in that it can be cast as a Sequential Second Order
Cone Programming problem yielding a polynomial time complexity.
The idea of learning from data locally and globally is also applicable in
regression tasks. Directly motivated from the Maxi-Min Margin Machine, a
new regression model named Local Support Vector Regression (LSVR) is
proposed in this book. LSVR is demonstrated to provide a systematic and
automatic scheme to locally and exibly adapt the margin which is globally
xed in the standard Support Vector Regression (SVR), a state-of-the-art
regression model. Therefore, it can tolerate the noise adaptively. The proposed LSVR is promising in the sense that it not only captures the local
information of the data in approximating functions, but more importantly,
includes special cases, which enjoy a physical meaning very much similar to
the standard SVR. Both theoretical and empirical investigations demonstrate
the advantages of this new model.
Besides the above two important models, another important contribution
of this book is that we also develop a novel global learning model called
Minimum Error Minimax Probability Machine (MEMPM). Although still
within the framework of global learning, this model does not need to assume
any specic distribution beforehand and represents a distribution-free Bayes
optimal classier in a worst-case scenario. This thus makes the model distinguished from the traditional global learning models, especially the traditional
Bayes optimal classier. One promising feature of MEMPM is that it can
derive an explicit accuracy bound under a mild condition, leading to a good
generalization performance for future data.
The fourth contribution of this book is that we develop the Biased Minimax Probability Machine (BMPM) model. Even though it is a special case of
MEMPM, we highlight this model because BMPM provides the rst systematic and rigorous approach for a kind of important learning tasks, namely, the
biased learning or imbalanced learning. Dierent from traditional imbalanced
(biased) learning methods, BMPM can quantitatively and explicitly incorporate a bias for one class and consequently emphasize the more important
8.2 Future Work
163
class. A series of experiments demonstrate that BMPM is very promising in

imbalanced learning and medical diagnosis.
8.2 Future Work

The models developed in this book bridge the gap between local learning and
global learning. This brings a new viewpoint for both existing local models
and global models. Following the viewpoint of learning from data both globally and locally, there seems to be a lot of immediate directions both inside
and beyond the proposed models in this book.
8.2.1 Inside the Proposed Models
There are certainly a lot of work for improving the proposed models in this
book.
First, all the models proposed in this book including Minimum Error
Minimax Probability Machine, Maxi-Min Margin Machine and Local Support Vector Machine, involve in solving either a single Second Order Cone
Programming or a Sequential Second Order Cone Programming problem.
Although many optimization programs have demonstrated their good performance and mathematic tractability in solving this kind of problems, they
are designed for general purposes and may not adequately exploit the specic properties in our models. Therefore, it is highly possible and valuable to
develop some special optimization algorithms for speeding up their training.
In particular, Maxi-Min Margin Machine enjoys the feature of sparsity. By
taking advantages of this property, researchers have developed fast optimization algorithms for Support Vector Machine. It is therefore very interesting to
investigate whether similar procedures can be applied here. This interesting
topic deserves much attention and remains to be an open problem.
Second, an immediate problem for Minimum Error Minimax Probability
Machine is the possible presence of local optimum in the practical optimization procedures. While empirical evidence shows that the global optimum
can be attained in most of cases, the local optimum may occur when two
types of data are not well-separated. Conventional simulated annealing [6, 14]
or deterministic annealing methods [11, 12] are certainly possible ways to
attack this problem, however a formal approach that is either a regularization
augment or an algorithmic approximation may be proved more appropriate.
Third, as shown in this book, all the proposed models apply the kernelization trick to extend their applications to nonlinear tasks. However, it
is well known that some global information, e.g. the structure information,
may not be well kept when the data are mapped from the original space to
the feature space. This may restrict the power of learning from data both
globally and locally. Motivated from this view, it is thus highly valuable to
develop techniques to retain the global information of data when performing
164
References
the projection from the original space to the feature space. This can also
be considered as a task on how to choose a suitable kernel, which currently
attracts much interest in the machine learning community [4, 15].
Another important future direction for the proposed classication models,
i.e. Minimum Error Minimax Probability Machine and Maxi-Min Margin
Machine, is how to extend the current binary classications into multi-way
classications. Although one vs. all and one vs. one [1, 16] approaches present
the main tools for conducting the upgrading, one always prefers to a more
systematic and more rigorous approach.
8.2.2 Beyond the Proposed Models
Although several important models have been motivated and developed from
the viewpoint of learning from data both globally and locally, beyond these
models there are plenty of work deserving future investigations.
One natural question is whether other famous local models or global models can be extended by engaging the viewpoint of learning from data globally
and locally. For example, Neural Networks, a large family of popular learning
models, might be also considered as modelling data in a local fashion. It is
therefore very interesting to investigate whether global information can also
be incorporated into these kinds of learning processes.
It is noted that the learning discussed in this book is restricted within
the framework of either classication or regression tasks. Both tasks belong
to the so-called supervised learning [5, 9, 18]. However, the other largely
dierent learning paradigm, unsupervised learning [10, 13, 17], and the recently emerging semi-supervised learning [2, 3, 8, 7] are not considered. Therefore, exploring possible applications of hybrid learning in this eld presents
a straightforward and immediate ongoing topic.
References
1. Allwein EL, Schapire RE, Singer Y (2000) Reducing multiclass to binary: A
unifying approach for margin classiers. Journal of Machine Learning Research
1:113141
2. Altun Y, McAllester D, Belkin M (2005) Maximum margin semi-supervised
learning for structured variables. In Advances in Neural Information Processing
Systerm (NIPS 18)
3. Ando R, Zhang T (2005) A framework for learning predictive structures from
multiple tasks and unlabeled data. Journal of Machine Learning Research
6:1817C1853
4. Bach FR, Lanckriet GRG, Jordan MI (2004) Multiple kernel learning, conic
duality, and the SMO algorithm. In Proceedings of International Conference
on Machine Learning (ICML-2004)
5. Bartlett PL (1998) Learning theory and generalization for neural networks
and other supervised learning techniques. In Neural Information Processing
Systems Tutorial
References
165
6. Cerny V (1985) Thermodynamical approach to the traveling salesman problem:

An ecient simulation algorithm. J. Opt. Theory Appl. 45(1):4151
7. Chapelle O, Zien A, Scholkopf B (2006) Semi-supervised learning. Cambridge,
MA: The MIT Press
8. Chawla NV, Karakoulas G (2005) Learning from labeled and unlabeled data:
An empirical study across techniques and domains. Journal of Articial Intelligence Research 23:331C366
9. Dietterich TG (1997) Machine learning research: Four current directions. AI
Magazine 18(4):97136
10. Dougherty James, Kohavi Ron, Sahami Mehran (1995) Supervised and unsupervised discretization of continuous features. In International Conference on
Machine Learning 194202
11. Dueck G (1993) New optimization heuristics :the great deluge algorithm and
the record-to-record travel. Journal of Computational Physics 104:8692
12. Dueck G, Scheurer T (1990) Threshold accepting: A general purpose optimization algorithm. Journal of Computational Physics 90:161175
13. Figueiredo M, Jain AK (2002) Unsupervised learning of nite mixture models.
Transaction on Pattern Analysis and Machine Intelligence 24(3):381396
14. Kirkpatrick S, Gelatt Jr CD, Vecchi MP (1983) Optimization by simulated
annealing. Science 220:671680
15. Lanckriet GRG, Cristianini N, Ghaoui LEl, Bartlett PL, Jordan MI (2004)
Learning the kernel matrix with semidenite programming. Journal of Machine
Learning Research
16. Rifkin R, Klautau A (2004) In defense of one vs. all classication. Journal of
Machine Learning Research 5:101141
17. Steck H, Jaakkola T (2002) Unsupervised active learning in large domains. In
Proceedings of the Eighteenth Annual Conference on Uncertainty in Articial
Intelligence
18. Wettig Hannes, Grunwald Peter, Roos Teemu (2002) Supervised naive Bayes
parameters. In Alasiuru P, Kasko S, editors, The Art of Natural and Articial:
Proceedings of the 10th Finnish Articial Intelligence Conference 7283
Index
A
AutoRegression (AR) 143, 147
B
Bayes optimal Hyperplane 33, 38
Bayesian Average Learning 19
Bayesian Optimal Decision 2
Bayes Point Machine 19
Bayesian Networks 1
Biased Classication 33
Biased Minimax Probability Machine
(BMPM) 33, 97
C
C4.5 105
Central Limit Theorem 40
Conic Programming 70
Concave-convex FP 36
Conjugate Gradient method 36
Cross validations 91
D
Data Orientation 76
Data Scattering Magnitude 76
Deterministic Annealing 161
Dictionary 127
Distribution-free 32
Divide and Conquer 73
Down-sampling 98
Down Side Mean Absolute
Error (DMAE) 142
E
Expectation Maximization (EM) 19
F
Financial time series 129
Fisher Discriminant Analysis (FDA)
77
Fixed and Asymmetrical Margin
(FAAM) 137
Fixed and Symmetrical Margin
(FASM) 136
Fractional Programming (FP) 36
G
Gabriel Graph 4
Game Theory 32
Gaussian Mixture Models 1
Generalized AutoRegressive Conditionally Heteroscedastic (GARCH)
168
Index
141
Generative Learning 16
Global Learning 16
Global Modeling 1
H
Hidden Markov Models 1
Hybrid Learning 5, 24
I
Maximum Geometry Mean (MGM)

100
Maximum Likelihood (ML) 17
Maximum Sum (MS) 100, 101
Mean Absolute Error (MAE) 141
Mercers Theorem 125, 136
Minimax Probability Machine
(MPM) 31
Minimum Cost (MC) 100
Minimum Error Minimax Probability
Machine (MEMPM) 21, 29
Momentum 139, 143
Imbalanced Learning 97
Independent, Identically Distribution- N
al (i.i.d.) 18
Naive Bayesian (NB) 16, 102
Non-xed and Symmetrical Margin
K
(NASM) 137
Non-xed and Asymmetrical Margin
k-Nearest-Neighbor 19,20,105
(NAAM) 137
Kernelization 45, 84, 125
Non-parametric Learning 19
Nonseparable Case 79
L
Lagrangian Multiplier 34
Large margin classiers 22, 69
Line Search 38
Locally and Globally 69
Local Modeling 3
Local Learning 22
Local Support Vector
Regression (LSVR) 119, 121
lpp-SVM 72
Lyapunov Condition 40
O
Over-tting 23
P
Parametric Method 41
Parzen Window 19, 20
Pseudo-concave Problem 36
Q
M
Mahalanobis Distance 72
Markov Chain Monte Carlo 19
Marshall and Olkin Theory 30
Maxi-Min Margin Machine (M4) 6,
25, 69
Maximum A Posterior (MAP) 17
Maximum Conditional Learning 18
Maximum Entropy Estimation 19
Quadratic Interpolation (QI) 38

Quadratic Programming (QP) 134
R
RBF Network 148
Receiver Operating Characteristic
(ROC) 100, 102
Recidivism 105
Index
Reduction 83
Robust Version 43
Rooftop 107
Rosen gradient projection 36
169
T
Tikhonovs Variation Method 80
Second Order Cone Programming

(SOCP) 70, 73,125
Sedumi 74
Sensitivity 111
Separable Case 71
Sequential Biased Minimax Probability Machine (SBMPM) 34, 36
Sequential Minimal Optimization 93
Simulated Annealing 161
Sinc Data 128
Sparse Approximation 127
Specicity 111
Statistical Learning 7
Structural Risk Minimization (SRM)
23, 134
Supervised Learning 162
Support Vector 70
Support Vector Machine 5, 22
Support Vector Regression (SVR)
119, 122, 134
Unbiased classication 33
Unsupervised Learning 162
Up-sampling 98
Up Side Mean Absolute Error
(UMAE) 142
V
Variational Margin Setting 134
VC dimension 24
Vector Recovery Index 65, 100
v-SVR 136
W
Weighted Support Vector Machine
34
Worst-case 32, 38
(n; k; )-bound problem 57

Huang - Machine Learning Modeling Data Locally and Globally - 2008

Enviado por

Dados do documento

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Huang - Machine Learning Modeling Data Locally and Globally - 2008

Enviado por

Direitos autorais:

Formatos disponíveis

ADVANCED TOPICS

IN SCIENCE AND TECHNOLOGY IN CHINA

Dr. Haiqin Yang,

Prof. Irwin King,

Dr. Michael Lyu,

ISBN 978-7-308-05831-5 Zhejiang University Press, Hangzhou

Machine Learning: Modeling Data Locally and Globally delivers the

Global Learning vs. Local Learning . . . . . . . . . . . . . . . . . . . . . . .

A General Global Learning Model: MEMPM . . . . . . . . . . . . .

3.2.6 Geometrical Interpretation . . . . . . . . . . . . . . . . . . . . . . . .

Learning Locally and Globally: Maxi-Min Margin Machine

Extension I: BMPM for Imbalanced Learning . . . . . . . . . . . .

5.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Extension II: A Regression Model from M4 . . . . . . . . . . . . . . .

Extension III: Variational Margin Settings within Local

Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

8.2.2 Beyond the Proposed Models . . . . . . . . . . . . . . . . . . . . . . 164

The objective of this book is to establish a framework which combines two

1.1 Learning and Global Modeling

Fig. 1.1. Two classes of two-dimensional data

1.2 Learning and Local Modeling

Fig. 1.2. An illustration of distribution-based classications (also known as

1.2 Learning and Local Modeling

Fig. 1.3. An illustration of local learning (also known as the Gabriel

1.4 Major Contributions

More specically, the class associated with s is obviously more scattered

1.3 Hybrid Learning

1.4 Major Contributions

A unied framework of many important models

1.4 Major Contributions

Containing explicit generalization bounds for both classes of data

1.6 Book Organization

10. Girosi F (1998) An equivalence between sparse approximation and support

2 Global Learning vs. Local Learning

Fig. 2.1. A hierarchy graph of statistical learning models

from the traditional global generative learning. As a critical contribution,

2.1 Problem Denition

2.1 Problem Denition

2 Global Learning vs. Local Learning

2.2 Global Learning

Depending on how the model is assumed on D, global learning can

2.2 Global Learning

2.2.1.1 Maximum Likelihood Learning & Maximum A Posterior

while in ML, the parameters are given as:

log p(z j |). (2.9)

2 Global Learning vs. Local Learning

2.2.1.2 Maximum Conditional Learning

where C = {c1 , c2 , . . . , cN } is the vector formed by the class label of each

By maximizing the conditional likelihood, MC is thus more direct and

2.2 Global Learning

where hN is dened as the length of the edge of RN .

2 Global Learning vs. Local Learning

Fig. 2.2. An illustration of Parzen Window estimation

Further, the posterior probability can be calculated as below:

2.2 Global Learning

2 Global Learning vs. Local Learning

2.3 Local Learning

2.4 Hybrid Learning

Fig. 2.3. An illustration of Support Vector Machine

2.4 Hybrid Learning

2 Global Learning vs. Local Learning

(represented by s) or the short axis of the x data (represented by s).

Fig. 2.4. A decision hyperplane with considerations of

2.5 Maxi-Min Margin Machine

By dening := 1/ we rewrite the dual problem as:

: x x 1/2 u = y + (0 ) y 1/2 v, u2 , v2 . (3.31)